[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_489":490,"_491":492},"routes/_app.x.$handle.$name",{"_99":98,"_32":44,"_107":89,"_26":108,"_13":14,"_109":-5,"_11":110,"_111":112,"_113":114,"_115":116,"_117":118,"_20":-5,"_119":-5,"_120":-5,"_121":24,"_122":123,"_124":58,"_125":-5,"_126":-5,"_127":128},"routes/_app.x.$handle.$name.branch.$branch.version.$version.diff",{"_11":12,"_13":14,"_15":16,"_17":-5,"_18":-5},"project",{"_42":43,"_32":44,"_37":45,"_46":-5,"_47":24,"_48":49,"_50":49,"_51":52,"_53":-5,"_18":54,"_55":56,"_57":58,"_59":58,"_60":-5,"_61":62,"_63":64,"_65":66,"_67":68,"_69":70,"_71":62,"_72":73,"_74":73,"_75":76,"_77":76,"_78":76,"_79":73,"_80":73,"_81":82,"_83":84,"_85":86,"_87":24},"version",1088,"changes",[19],"aiChatInfo","userId",{"_20":21,"_22":-7,"_23":24,"_25":24,"_26":27,"_28":29},"file",{"_30":31,"_32":33,"_34":20,"_35":36,"_37":38,"_39":14,"_40":41,"_23":24},"previous","deleted",false,"deletedHere","path","posts/phoenix-copy.md","latestPath","posts/learn-to-code.md","fileId","3aeb30c6-797b-11f1-95e5-1607ee4eb77e","name","phoenix-copy.md","fileType","content","---\ntitle: On Chad Fowler's Phoenix Architecture\ntags: ai, coding\ndate: 2026-06-30\n---\n\n# On Chad Fowler's _Phoenix Architecture_\n\nJune 30,2026\n\nI kept hearing about Chad Fowler's\n[Phoenix Architecture](https://aicoding.leaflet.pub/), and after\n[Charity Majors' endorsement](https://charitydotwtf.substack.com/p/ai-demands-more-engineering-discipline),\nI figured it was worth an investigation.\n\nThe core premise: LLMs make code cheap to generate, so code should be\ndisposable.\n\nThe most compelling parts are the analogies to immutability:\n\n> Immutable infrastructure. Stateless services. Containers. Blue-green\n> deployments. Infrastructure as code.\n>\n> These ideas all share a common premise: never fix a running thing. Replace it.\n>\n> AI pushes this premise beyond infrastructure and into application code itself.\n> When rewriting is cheap, editing in place becomes risky. Mutation accumulates\n> entropy. Replacement resets it.\n\n> When you edit code in place, you're doing the software equivalent of SSHing\n> into a production server and tweaking a config file.\n\nI whole-heartedly agree that mutating something low-level is bad. Regenerating\nit from scratch from something high-level is good. This is how we moved from\nJQuery to ReactJS or from Binary to Assembly to C to JavaScript – higher level\nabstractions that let us _generate_ disposable lower level artifacts.\n\nBut that begs the question: regenerate from _what_? I sure hope you don't mean\nnatural language specs! Don't make me get the comic out again.\n\n![A four-panel CommitStrip comic. Panel 1: A man holding a coffee mug says to a colleague typing on a laptop, \"Some day we won't even need coders any more. We'll be able to just write the specification and the program will write itself.\" Panel 2: The colleague replies, \"Oh wow, you're right! We'll be able to write a comprehensive and precise spec and bam, we won't need programmers any more!\" The first man sips his coffee and says, \"Exactly!\" Panel 3: The first man asks, \"And do you know the industry term for a project specification that is comprehensive and precise enough to generate a program?\" The colleague answers, \"Uh... no....\" Panel 4: The colleague says, \"Code,\" and then, \"It's called code.\"](https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/f7e9e621-0fad-4c3a-a65b-d16832c02b00/public)\n\nI made this case at length in\n[Reports of code's death are greatly exaggerated](https://stevekrouse.com/precision),\nbut in brief: natural language is a terrible \"source of truth\" for software. As\nBertrand Russell says, \"Everything is vague to a degree you do not realize till\nyou have tried to make it precise.\" If today's code feels too low-level and\nboilerplate-y in a way that makes you want to burn it down and regenerate it\nfrom scratch, the fix is to climb incrementally higher to better programming\nabstractions that let you convey your higher-level intent while still being\nprecise and mastering the inherent complexity.\n\nIt's a category error to conclude that because code is cheap to generate, we\nshould treat it all as disposable. It's leaning a bit too far into the vibes of\n2025 and Gas Town, when designing around lots of slop seemed like the thing to\ndo. I don't know if we've yet hit peak code slop – probably not quite yet – but\nit feels like that's right around the corner.\n\nSo while I don't agree with the solutions of the Phoenix Architecture, I do\nagree with the problem:\n\n> The limiting factor is no longer writing software, but understanding,\n> evaluating, and governing it.\n\nPeter Naur's\n[Programming as Theory Building](https://pages.cs.wisc.edu/~remzi/Naur.pdf), has\na similar theme: the code doesn't represent very much of the \"theory\" that made\nit. However, Naur warns you against thinking you could put all that \"theory\"\ninto specs or docs. It lives in the heads of the creators.\n\nBut now that we have ephemeral machine \"heads\" – where does that understanding\nlive long-term? It kinda lives in the documentation that the machines write for\nthemselves, but at least for now, we all know that doesn't scale the same way\nhuman understanding does.\n\nFor improving code understanding, I always return to Glen Chiacchieri's\n[A Human-Readable Interactive Representation of a Code Library](https://glench.github.io/fuzzyset.js/ui/),\na delightful explorable explanation of a fuzzyset library. (If you care at all\nabout code understanding, please go play with it!) Now with vibe coding, there's\nvery little excuse to not generate artifacts like these for _every_ piece of\nsoftware we build.\n\nMy personal thesis is that making software understandable to humans also makes\nit understandable to LLMs, and vice versa, so investments in understandability\ndo double-duty. This is great news! We can keep humans and machines both\nin-the-loop at the same time. I think that's the path towards accelerating\ndevelopment, and making the most of how cheap code is now to generate.","createdAt",["D",1783370560154],"branchVersion","branchId","9724c258-e985-11ef-b2c8-e6cdfca9ef9f","id","972467e0-e985-11ef-b2c8-e6cdfca9ef9f","stevekrouse_dot_com",["D",1739394544742],"deletedAt","legacyEnvVars","privacy","public","httpPrivacy","description","The personal site of Steve Krouse, stevekrouse.com","imageUrl","a0bf3b31-15a5-4d5c-880e-4b1e22c9bc18","pinnedFileId","d8c12d22-b19f-11f0-847d-0224a6c84d84","pinPreview",true,"pinMigrated","screenshotVersion","likeCount",0,"orgId","e81e7335-6d4c-4764-a902-d0962bbfa9f7","tursoName","972467e0e98511efb2c8e6cdfca9ef9f","tursoHostname","972467e0e98511efb2c8e6cdfca9ef9f-val-town.aws-us-east-2.turso.io","tursoMeasuredAt",["D",1789559472742],"pinned","userPerm","r","valPerm","blobPerm","rw","emailPerm","sqlitePerm","projectPerm","telemetryPerm","user",{"_99":98,"_106":102},"org",{"_99":98,"_100":-5,"_101":102,"_18":54,"_103":104,"_105":24},"branch",{"_13":88,"_32":89,"_42":41,"_90":43,"_91":-5,"_92":-5,"_93":-5,"_94":-5,"_95":-5,"_96":44,"_97":98},"isOwned",1160,"main","projectId","forkedBranchId","forkedBranchVersion","forkedProjectPrivacy","forkedProjectId","forkedProjectOrgId","projectName","userHandle","stevekrouse","handle","displayName","avatar","https://images.clerk.dev/uploaded/img_2PqHa2Gsy93xQrjh2w78Xu0cChW.jpeg","tier","pro","canRestrictVals","profileImageUrl","branchName","","adoptOrgId",{"_42":43,"_32":44,"_37":484,"_46":-5,"_47":24,"_48":49,"_50":49,"_51":52,"_53":-5,"_18":54,"_55":56,"_57":58,"_59":58,"_60":-5,"_61":62,"_63":64,"_65":66,"_67":68,"_69":485,"_71":62,"_72":73,"_74":73,"_75":76,"_77":76,"_78":76,"_79":73,"_80":73,"_81":486,"_83":487,"_85":488,"_87":24},"tree",[250,271,272,273,274,275,276,277,278,279,280,281],"files",[250,251,252,253,254,255,256,257,258,259,260,261,262,263,264,265,266,267,268,269,270,271,272,273,274,275,276,277,278,279,280,281],"counts",{"_239":62,"_240":206,"_241":242,"_243":244,"_245":246},"contents",[130,131,132,133,134,135,136,137,138,139,140,141],"upstreamProject","upstreamBranch","emptyEnvVars","flags",[],"hasHttpFiles","grants","orgs","customDomains",{"_56":129},"stevekrouse.com",{"_42":232,"_32":233,"_30":234,"_145":-5,"_40":41,"_39":235,"_147":58,"_37":236,"_13":149,"_34":237,"_35":-5,"_23":24,"_152":-5,"_26":238},{"_42":224,"_32":225,"_30":226,"_145":-5,"_40":41,"_39":227,"_147":24,"_37":228,"_13":229,"_34":150,"_35":230,"_23":24,"_152":-5,"_26":231},{"_42":217,"_32":218,"_30":219,"_145":-5,"_40":41,"_39":220,"_147":58,"_37":221,"_13":62,"_34":150,"_35":222,"_23":24,"_152":-5,"_26":223},{"_42":209,"_32":210,"_30":211,"_145":-5,"_40":41,"_39":212,"_147":58,"_37":213,"_13":214,"_34":150,"_35":215,"_23":24,"_152":-5,"_26":216},{"_42":201,"_32":202,"_30":203,"_145":-5,"_40":41,"_39":204,"_147":58,"_37":205,"_13":206,"_34":20,"_35":207,"_23":24,"_152":-5,"_26":208},{"_42":193,"_32":194,"_30":195,"_145":-5,"_40":41,"_39":196,"_147":58,"_37":197,"_13":198,"_34":150,"_35":199,"_23":24,"_152":-5,"_26":200},{"_42":185,"_32":186,"_30":187,"_145":-5,"_40":41,"_39":188,"_147":58,"_37":189,"_13":190,"_34":150,"_35":191,"_23":24,"_152":-5,"_26":192},{"_42":177,"_32":178,"_30":179,"_145":-5,"_40":41,"_39":180,"_147":24,"_37":181,"_13":182,"_34":150,"_35":183,"_23":24,"_152":-5,"_26":184},{"_42":169,"_32":170,"_30":56,"_145":-5,"_40":41,"_39":171,"_147":24,"_37":172,"_13":173,"_34":174,"_35":175,"_23":24,"_152":98,"_26":176},{"_42":161,"_32":162,"_30":163,"_145":-5,"_40":41,"_39":164,"_147":24,"_37":165,"_13":166,"_34":150,"_35":167,"_23":24,"_152":-5,"_26":168},{"_42":154,"_32":155,"_30":156,"_145":-5,"_40":41,"_39":157,"_147":58,"_37":158,"_13":62,"_34":150,"_35":159,"_23":24,"_152":-5,"_26":160},{"_42":142,"_32":143,"_30":144,"_145":-5,"_40":41,"_39":146,"_147":58,"_37":148,"_13":149,"_34":150,"_35":151,"_23":24,"_152":-5,"_26":153},"1c38beb8-1a3b-11f1-93e1-42dde27851f2","rss.tsx","fd71afee-1a3a-11f1-a277-42dde27851f2","parentId",729,"latest",["D",1772897660393],1,"script","import { getPostMetadata, ensurePostsLoaded, type PostMeta } from \"./post-helpers.tsx\";\n\nconst SITE_URL = \"https://stevekrouse.com\";\n\nfunction escapeXml(str: string): string {\n return str\n .replace(/&/g, \"&\")\n .replace(//g, \">\")\n .replace(/\"/g, \""\")\n .replace(/'/g, \"'\");\n}\n\nexport async function generateRSSFeed(): Promise {\n await ensurePostsLoaded();\n\n const posts = getPostMetadata()\n .filter((p) => p.date)\n .sort((a, b) => new Date(b.date!).getTime() - new Date(a.date!).getTime());\n\n const items = posts\n .map(\n (p) => ` \n ${escapeXml(p.title)}\n ${SITE_URL}/${p.slug}\n ${SITE_URL}/${p.slug}\n ${new Date(p.date!).toUTCString()}${\n p.description\n ? `\\n ${escapeXml(p.description)}`\n : \"\"\n }\n `,\n )\n .join(\"\\n\");\n\n const xml = `\n\n \n Steve Krouse\n ${SITE_URL}\n Essays by Steve Krouse\n en-us\n \n${items}\n \n`;\n\n return new Response(xml, {\n headers: {\n \"Content-Type\": \"application/rss+xml; charset=utf-8\",\n },\n });\n}\n","label",[143],"785c9d3c-4672-11f1-95cb-42b51c65c3df","remark-autolink-bare.ts","785c9df0-4672-11f1-95cb-42b51c65c3df",953,["D",1777759288465],"// A small remark plugin that auto-linkifies bare URLs like\n// \"stevekrouse.com/abstraction\" or \"blog.val.town/vibe-code\" — i.e. URLs\n// without a protocol that GFM's autolink extension intentionally won't match.\n//\n// It walks the markdown AST, finds text nodes, and splits each one wherever\n// it sees a URL-looking pattern, replacing those slices with link nodes.\n//\n// We're intentionally conservative: we require either a path/query/fragment\n// after the domain (so `foo.com/bar`, `foo.com?x=1`) OR a known TLD, so we\n// don't accidentally turn \"Hello world.\" into a link.\n\nimport { visit } from \"https://esm.sh/unist-util-visit@5\";\n\n// Match: optional subdomain(s), domain, dot, TLD, then required path/query/fragment.\n// Stop at whitespace, common punctuation followed by space/end, or markdown markers.\n// Trailing punctuation (.,;:!?) is excluded so \"see foo.com/bar.\" doesn't include the period.\nconst BARE_URL_REGEX =\n /\\b([a-z0-9](?:[a-z0-9-]*[a-z0-9])?(?:\\.[a-z0-9](?:[a-z0-9-]*[a-z0-9])?)+\\/[^\\s<>()[\\]]*[^\\s<>()[\\]\\.,;:!?])/gi;\n\nexport default function remarkAutolinkBare() {\n return (tree: any) => {\n visit(tree, \"text\", (node: any, index: number | undefined, parent: any) => {\n if (!parent || index === undefined) return;\n // Don't linkify text inside existing links or code\n if (parent.type === \"link\" || parent.type === \"linkReference\") return;\n\n const value: string = node.value;\n const matches = [...value.matchAll(BARE_URL_REGEX)];\n if (matches.length === 0) return;\n\n const newNodes: any[] = [];\n let lastIndex = 0;\n\n for (const match of matches) {\n const url = match[0];\n const start = match.index!;\n if (start > lastIndex) {\n newNodes.push({ type: \"text\", value: value.slice(lastIndex, start) });\n }\n newNodes.push({\n type: \"link\",\n url: `https://${url}`,\n children: [{ type: \"text\", value: url }],\n });\n lastIndex = start + url.length;\n }\n if (lastIndex < value.length) {\n newNodes.push({ type: \"text\", value: value.slice(lastIndex) });\n }\n\n parent.children.splice(index, 1, ...newNodes);\n return index + newNodes.length;\n });\n };\n}\n",[155],"5962ed80-a801-494d-93e5-39f2a94889a4","post-helpers.tsx","68026ad8-b366-11f0-9979-0224a6c84d84",1082,["D",1782848955587],50,"/** @jsxImportSource https://esm.sh/react@18.2.0 */\nimport { renderToString } from \"https://esm.sh/react-dom@18.2.0/server\";\nimport {\n listFilesByPath,\n readFile,\n} from \"https://esm.town/v/std/utils/index.ts\";\nimport ReactMarkdown from \"https://esm.sh/react-markdown@9?deps=react@18.2.0\";\nimport rehypeSlug from \"https://esm.sh/rehype-slug@6?deps=react@18.2.0\";\nimport rehypeAutolinkHeadings from \"https://esm.sh/rehype-autolink-headings@7?deps=react@18.2.0\";\nimport rehypeRaw from \"https://esm.sh/rehype-raw@7?deps=react@18.2.0\";\nimport remarkGfm from \"https://esm.sh/remark-gfm@4?deps=react@18.2.0\";\nimport remarkAutolinkBare from \"./remark-autolink-bare.ts\";\nimport ValTown from \"npm:@valtown/sdk@1.9.0\";\nimport { PostTemplate } from \"./PostTemplate.tsx\";\nimport { NotFound } from \"./NotFound.tsx\";\n\nexport interface PostMeta {\n title: string;\n tags: string[];\n slug: string;\n date?: string;\n description?: string;\n}\n\nfunction parseFrontmatter(markdown: string): { meta: Record; content: string } {\n const match = markdown.match(/^---\\n([\\s\\S]*?)\\n---\\n([\\s\\S]*)$/);\n if (!match) return { meta: {}, content: markdown };\n const meta: Record = {};\n for (const line of match[1].split(\"\\n\")) {\n const idx = line.indexOf(\":\");\n if (idx > 0) {\n meta[line.slice(0, idx).trim()] = line.slice(idx + 1).trim();\n }\n }\n return { meta, content: match[2] };\n}\n\n// Shared metadata for all loaded posts\nlet postMetadata: PostMeta[] = [];\n\nexport function getPostMetadata(): PostMeta[] {\n return postMetadata;\n}\n\nasync function loadFile(file: ValTown.Vals.Files.FileRetrieveResponse) {\n if (file.name.endsWith(\".tsx\")) {\n return (await import(file.links.module)).default;\n } else if (file.name.endsWith(\".html\")) {\n const html = await readFile(file.path);\n return () =>\n new Response(html, { headers: { \"Content-Type\": \"text/html\" } });\n } else if (file.name.endsWith(\".md\")) {\n const raw = await readFile(file.path);\n const { meta, content } = parseFrontmatter(raw);\n\n // Title from frontmatter, or first h1, or filename\n const titleMatch = content.match(/^#\\s+(.+)$/m);\n const title = meta.title || (titleMatch ? titleMatch[1] : file.name.replace(\".md\", \"\"));\n\n // Tags from frontmatter (comma-separated)\n const tags = meta.tags\n ? meta.tags.split(\",\").map((t: string) => t.trim()).filter(Boolean)\n : [];\n\n const date = meta.date || undefined;\n const description = meta.description || undefined;\n\n // OG image: explicit frontmatter, or auto-detect first image in markdown\n const firstImageMatch = content.match(/!\\[.*?\\]\$(.*?)\$/);\n const ogImage = meta.ogImage || (firstImageMatch ? firstImageMatch[1] : undefined);\n\n const slug = file.name.split(\".\")[0].toLowerCase();\n\n // Store metadata\n postMetadata.push({ title, tags, slug, date, description });\n\n return function () {\n return new Response(\n renderToString(\n \n {content}\n ,\n ),\n { headers: { \"Content-Type\": \"text/html\" } },\n );\n };\n }\n}\n\nasync function getPosts() {\n postMetadata = []; // reset on reload\n const posts = Object.fromEntries(\n await Promise.all(\n (await listFilesByPath(\"posts\"))\n .map(\n async (\n file,\n ) => [\n file.name.split(\".\")[0].toLowerCase(),\n await loadFile(file), // could defer this until someone wants this particular file\n ],\n ),\n ),\n );\n return posts;\n}\nasync function getPostsBounded() {\n const posts = await Promise.race([\n getPosts(),\n new Promise((resolve) => setTimeout(resolve, 5000)),\n ]);\n if (!posts) {\n console.error(\"Timeout fetching posts\");\n return [];\n }\n console.log(`Fetched ${Object.keys(posts).length} posts`);\n return posts;\n}\n\nlet posts;\nexport async function ensurePostsLoaded() {\n if (!posts) posts = await getPostsBounded();\n}\nexport async function renderPage(req: Request) {\n await ensurePostsLoaded();\n console.log(posts);\n\n // render /posts/* in this val\n const url = new URL(req.url);\n const name = url.pathname.slice(1);\n const post = posts[name.toLowerCase()];\n if (post) return post(req);\n\n // look up page on github pages\n const githubPagesUrl =\n `https://stevekrouse.github.io${url.pathname}${url.search}`;\n const response = await fetch(\n new Request(githubPagesUrl, {\n method: req.method,\n headers: new Headers({\n ...Object.fromEntries(req.headers),\n \"User-Agent\": \"Val Town\",\n }),\n }),\n );\n if (response.status !== 404) return response;\n\n // return 404\n return NotFound(url.pathname + url.search);\n}\n",[162],"a67879b1-bc28-4526-8375-a010a6ca234d","main.tsx",995,["D",1781346741270],43,"http","/** @jsxImportSource https://esm.sh/react@18.2.0 */\nimport { renderToString } from \"https://esm.sh/react-dom@18.2.0/server\";\nimport { Homepage } from \"./index.tsx\";\nimport { Hono } from \"npm:hono@4.0.0\";\nimport {\n ensurePostsLoaded,\n getPostMetadata,\n renderPage,\n} from \"./post-helpers.tsx\";\nimport { generateRSSFeed } from \"./rss.tsx\";\nimport { listFilesByPath } from \"https://esm.town/v/std/utils/index.ts\";\nimport { PostTemplate } from \"./PostTemplate.tsx\";\nimport { hotReload } from \"https://esm.town/v/std/utils/index.ts\";\n\nconst SOURCE = \"https://www.val.town/x/stevekrouse/stevekrouse_dot_com\";\n\n// Cache blog posts metadata for source links\nlet blogPostsCache: Map | null = null;\n\nasync function getBlogPostsMap() {\n if (blogPostsCache) return blogPostsCache;\n\n const files = await listFilesByPath(\"posts\");\n blogPostsCache = new Map(\n files.map((file) => {\n const slug = file.name.split(\".\")[0].toLowerCase();\n const sourceUrl = `${SOURCE}/code/${file.path}`;\n return [slug, sourceUrl];\n }),\n );\n\n return blogPostsCache;\n}\n\nfunction getBlogPostBySlug(slug: string, postsMap: Map) {\n const sourceUrl = postsMap.get(slug);\n if (!sourceUrl) return null;\n return { valSourceURL: sourceUrl };\n}\n\nasync function renderTagsPage(tag?: string) {\n await ensurePostsLoaded();\n const allPosts = getPostMetadata();\n\n // Build tag -> posts map\n const tagMap = new Map();\n for (const post of allPosts) {\n for (const t of post.tags) {\n if (!tagMap.has(t)) tagMap.set(t, []);\n tagMap.get(t)!.push({ title: post.title, slug: post.slug });\n }\n }\n\n // Sort tags alphabetically\n const sortedTags = [...tagMap.keys()].sort((a, b) => a.localeCompare(b));\n\n if (tag) {\n // Single tag page\n const decodedTag = decodeURIComponent(tag);\n const postsForTag = tagMap.get(decodedTag) || [];\n return new Response(\n renderToString(\n \n

Posts tagged \"{decodedTag}\"

\n ← All tags\n

\n {postsForTag.length === 0\n ?

No posts found with this tag.

\n : (\n

\n {p.title}\n

\n )}\n ,\n ),\n { headers: { \"Content-Type\": \"text/html\" } },\n );\n }\n\n // All tags page\n return new Response(\n renderToString(\n \n

\n Steve Krouse\n

\n 👋 Hi, I'm Steve. I live in Prospect Heights, Brooklyn.\n

\n I build{\" \"}\n Val Town, a delightful way to\n write and deploy code in the cloud.\n

\n I founded the Future of Coding podcast & community and The Coding\n Space after-school program. I've worked at Looker and First Round\n Capital. I studied computer science at Penn.\n

\n We're hiring{\" \"}\n at Val Town. To learn how to best work with me, read my{\" \"}\n README.\n

\n This site was{\" \"}\n \n built in Val Town\n .\n

Projects

\n 2022 Jul -\n Val Town\n
\n 2022 Sep -\n Date Me Directory\n
\n 2022 Jan -\n \n Zaplib\n \n
\n 2021 Mar -\n Compose\n
\n 2017 Jul -\n \n Future of Coding\n \n
\n 2016 May -\n \n WoofJS\n \n
\n 2015 Jul -\n The Coding Space\n

\n\n

Essays

\n 2026 Jun -\n \n On Chad Fowler's Phoenix Architecture\n \n
\n 2026 May -\n \n AI Agents are briefly overhyped\n \n
\n 2026 Mar -\n \n steve-eval – getting AI to write like me\n \n
\n 2026 Mar -\n \n Reports of code's death are greatly exaggerated\n \n
\n 2026 Mar -\n \n What if you never had to get an API key ever again?\n \n
\n 2026 Mar -\n \n Personal Investor Relations\n \n
\n 2026 Feb -\n \n Intelligence Buying Intelligence\n \n
\n 2026 Jan -\n \n Advice to college students in 2026\n \n
\n 2025 Mar -\n \n Find a model\n \n
\n 2025 Feb -\n \n I proposed to my girlfriend in Quordle\n \n
\n 2022 Nov -\n \n End-programmer Programming\n \n
\n 2019 Apr -\n \n Lunch with Alan Kay: how to become educated enough to invent\n the future\n \n
\n 2018 Oct -\n \n The \"Next Big Thing\" is a Room\n \n
\n 2017 Oct -\n Life Plan\n
\n 2017 Oct -\n \n WoofJS.com One Year Later\n \n
\n 2017 Jul -\n \n The Coding Space: Two Year Update\n \n
\n 2017 Jul -\n \n Learn to code in two-player mode\n \n
\n 2017 May -\n \n The Rose Project (April 2017)\n \n
\n 2017 Mar -\n \n The Rose Project (March 2017)\n \n
\n 2017 Jan -\n \n Resources for creating front-end programming languages and\n frameworks in 2017\n \n
\n 2016 Nov -\n \n Types are shapes — a graphical programming exploration\n \n
\n 2016 Nov -\n \n TEDD — Technologically-Enabled Direct Democracy\n \n
\n 2016 Aug -\n \n Goodbye, Seymour\n \n
\n 2016 Jul -\n \n The Trick to Teach Anonymous Functions to 11-Year-Olds\n \n
\n 2016 Jul -\n \n Scratch Has a Marketing Problem\n \n
\n 2016 Jul -\n \n WoofJS — making JavaScript learnable\n \n
\n 2016 Jun -\n \n The Coding Space: A year in review\n \n
\n 2016 Feb -\n \n The 4 Steps to Giving Magical Praise\n \n
\n 2016 Jan -\n \n \"Why do we need to learn this?\"\n \n
\n 2015 Dec -\n The Lens Lens\n
\n 2015 Dec -\n \n The College of the Future\n \n
\n 2015 Dec -\n \n Stop Paying $250k for a Degree and a Network\n \n
\n 2015 Nov -\n \n Incidentally for machines\n \n
\n 2015 Jun -\n \n A Cursory Education On Education\n \n
\n 2014 Jun -\n \n One year ago today I blogged that I'd never drop out of\n college\n \n
\n 2013 Sep -\n \n Making friends at the world's largest hackathon\n \n
\n 2013 Jul -\n Be a quitter\n
\n 2013 Jun -\n \n What It Really Means To Drop Out\n \n
\n 2013 May -\n Paying it Forward\n
\n 2013 Mar -\n Being Nice Matters\n
\n 2013 Nov -\n \n When nobody's watching\n \n
\n 2013 Nov -\n \n Stuffed Animal Priorities\n \n
\n 2013 Feb -\n \n Medieval Luxury, Modern Torture, and Education 2.0\n \n
\n 2013 Jan -\n Theseus and The Minotaur\n
\n 2012 Dec -\n \n IP And Innovation: A Software Case Study\n \n
\n 2012 Dec -\n \n Scheming Sororities and the Gale-Shapley Algorithm\n \n
\n 2012 Dec -\n \n The Arbitrary and The Elegant\n \n
\n 2012 Nov -\n Start Small\n

\n\n

Podcasts

\n 2023 Aug -\n \n JSParty × val.town with Steve Krouse\n \n
\n 2023 Aug -\n \n Scaling Devtools × val.town with Steve Krouse\n \n
\n 2023 Jul -\n \n Supper Club × val.town with Steve Krouse\n \n
\n 2023 Jul -\n \n CS Primer – What makes programming fun?\n \n
\n 2023 Apr -\n \n Devtools.fm – val.town\n \n
\n 2017 Jul -\n \n Future of Coding Podcast\n \n
\n 2016 Oct -\n \n Off Trail Learning - Dropping Out of College & the Silicon\n Valley Dream\n \n

\n\n

\n Twitter |{\" \"}\n \n Bluesky\n {\" \"}\n | Github |{\" \"}\n LinkedIn |\n {\" \"}\n RSS\n

\n \n ,\n ),\n {\n headers: {\n \"Content-Type\": \"text/html\",\n // TODO only when on main - \"Cache-Control\": \"public, max-age=3600, stale-while-revalidate=0\",\n },\n },\n );\n}",[178],"ac7f5c48-b366-11f0-9268-0224a6c84d84","helpers.tsx","1fc9543e-efda-11ef-970d-e6cdfca9ef9f",477,["D",1761591401040],28,"/** @jsxImportSource https://esm.sh/react@18.2.0 */\nconst linkClass = \"text-blue-500 hover:underline\";\nexport const Link = (\n { children, href }: {\n children?: React.ReactNode;\n href: string;\n },\n) => {children};",[186],"c5cc456c-b366-11f0-9268-0224a6c84d84","analytics.ts","feaa60fe-e985-11ef-b2e0-e6cdfca9ef9f",479,["D",1761591443489],8,"import { sqlite } from \"https://esm.town/v/std/sqlite\";\n\nexport async function getHits() {\n const winner = await Promise.race([\n _getHits(),\n new Promise((resolve) => setTimeout(resolve, 1000)),\n ]);\n if (!winner) throw new Error(\"Took longer than 1 second to get hits\");\n return winner;\n}\n\nasync function _getHits() {\n const [, { rows: [[monthHits]] }, { rows: [[todayHits]] }] = await sqlite\n .batch([\n \"INSERT INTO stevekrouse_com_hits DEFAULT VALUES\",\n \"SELECT COUNT(*) FROM stevekrouse_com_hits where strftime('%m', timestamp) = strftime('%m', 'now') AND strftime('%Y', timestamp) = strftime('%Y', 'now')\",\n \"SELECT COUNT(*) from stevekrouse_com_hits where strftime('%m', timestamp) = strftime('%m', 'now') AND strftime('%Y', timestamp) = strftime('%Y', 'now') AND strftime('%d', timestamp) = strftime('%d', 'now')\",\n ]);\n return { monthHits, todayHits };\n}",[194],"ea88eeda-090d-11f1-bc78-42dde27851f2","README.md","c0b19768-e985-11ef-9203-e6cdfca9ef9f",631,["D",1771009079911],5,"# stevekrouse.com - my personal website\n\nThis project hosts my personal website.\n\nIt used to [live on Github Pages](https://github.com/stevekrouse/stevekrouse.github.io), which is why I proxy over requests to certain blog posts over to the Github Pages site still.\n\n## Architecture\n\n```mermaid\ngraph TD\n A[Incoming Request] --> B[\"main.tsx (Hono Router)\"]\n\n B -->|\"GET /\"| C[\"index.tsx → Homepage()\"]\n C --> D[\"analytics.ts → getHits()\"]\n D --> E[(SQLite DB)]\n C --> F[Server-side rendered HTML with React]\n\n B -->|\"GET /source\"| G[Redirect to Val Town source code]\n\n B -->|\"GET /*\"| H[\"post-helpers.tsx → renderPage()\"]\n\n H --> I{Match slug in posts/?}\n I -->|Yes .tsx| J[Dynamic import → run as handler]\n I -->|Yes .md| K[Parse Markdown with react-markdown]\n I -->|Yes .html| L[Serve raw HTML]\n J --> M[\"PostTemplate.tsx (layout wrapper)\"]\n K --> M\n I -->|No match| N[Proxy to GitHub Pages]\n N --> O{GitHub Pages 404?}\n O -->|No| P[Return GitHub Pages response]\n O -->|Yes| Q[Return 404 page]\n\n style B fill:#f97316,color:#fff\n style E fill:#3b82f6,color:#fff\n style M fill:#10b981,color:#fff\n```\n\n### Key Files\n\n| File | Role |\n|------|------|\n| `main.tsx` | HTTP entry point — Hono router handling `/`, `/source`, and `/*` |\n| `index.tsx` | Renders the homepage with projects, essays, podcasts, and analytics |\n| `analytics.ts` | Tracks page hits in SQLite; returns today & monthly counts |\n| `post-helpers.tsx` | Loads posts from `posts/` dir (.tsx, .md, .html), falls back to GitHub Pages |\n| `PostTemplate.tsx` | Shared blog post layout with OG tags, water.css, and nav header |\n| `helpers.tsx` | Shared `` component |\n| `posts/` | Blog posts in .tsx, .md, or .html format |\n\n## Todos \n\n- [ ] Speed up page load by loading sqlite data later like in @healeycodes/steve_web\n- [ ] Store more (legally storable) analytics data, and maybe make a sparkline!\n- [ ] Add some sort of way to contact me\n- [ ] Move over all my blog posts from Github Pages (maybe into @std/blob as a CMS?)",[202],"5252d2cb-5d43-4b0d-ae0d-bb77376c9751","PostTemplate.tsx","527e932c-f947-11f0-9fb8-42dde27851f2",1061,["D",1782848233329],37,"/** @jsxImportSource https://esm.sh/react@18.2.0 */\nimport { Footer } from \"./Footer.tsx\";\n\ninterface PostTemplateProps {\n title?: string;\n description?: string;\n ogImage?: string;\n tags?: string[];\n children: React.ReactNode;\n}\n\n/**\n * Template for blog posts with:\n * - Good OG meta tags\n * - Favicon (SK in orange)\n * - Header with link back to homepage\n * - View source link at the top\n * - Water.css styling\n */\nexport function PostTemplate({\n title = \"Steve Krouse\",\n description,\n ogImage,\n tags,\n children,\n}: PostTemplateProps) {\n // Generate favicon SVG (orange SK, transparent background)\n const faviconSvg = `data:image/svg+xml,${\n encodeURIComponent(`\n \n `)\n }`;\n\n return (\n \n \n {title}\n \n \n \n \n\n {/* OpenGraph meta tags */}\n \n \n {description && (\n <>\n \n \n \n \n )}\n {ogImage && (\n <>\n \n \n \n \n )}\n \n\n {/* Highlight.js for syntax highlighting (light + dark themes) */}\n \n \n \n \n\n {/* Mermaid for diagrams */}\n \n\n \n \n \n

\n 404\n

\n This page wandered off.\n

\n We couldn't find{\" \"}\n \n stevekrouse.com{path}\n \n

\n Maybe it's in a blog post somewhere. Try heading back home?\n

\n \n ← Back to stevekrouse.com\n \n

\n \n ,\n ),\n { status: 404, headers: { \"Content-Type\": \"text/html\" } },\n );\n}\n",[218],"3c14c978-9346-4c1f-97d4-6a60761d7814","Footer.tsx","83491854-2f4a-11f1-a46b-42dde27851f2",1036,["D",1782844059093],7,"/** @jsxImportSource https://esm.sh/react@18.2.0 */\n\nexport function Footer({ tags }: { tags?: string[] }) {\n return (\n \n );\n}",[225],"722a9b58-ed59-11ef-b30b-e6cdfca9ef9f","posts","6f89aa88-ed59-11ef-a745-e6cdfca9ef9f",42,["D",1739815389496],"directory",[233],"pulls","branches","forks",4,"envVars",{"_247":62,"_249":62},"connections",{"_247":62,"_248":62},"count","unfilledCount","emptyCount",{"_42":232,"_32":233,"_30":234,"_145":-5,"_40":41,"_39":235,"_147":58,"_37":481,"_13":149,"_34":237,"_35":-5,"_23":24,"_152":-5,"_283":482,"_26":483},{"_42":472,"_32":473,"_30":474,"_145":234,"_40":41,"_39":475,"_147":24,"_37":476,"_13":477,"_34":20,"_35":478,"_23":24,"_152":-5,"_283":479,"_26":480},{"_42":463,"_32":464,"_30":465,"_145":234,"_40":41,"_39":466,"_147":58,"_37":467,"_13":468,"_34":20,"_35":469,"_23":24,"_152":-5,"_283":470,"_26":471},{"_42":454,"_32":455,"_30":456,"_145":234,"_40":41,"_39":457,"_147":58,"_37":458,"_13":459,"_34":20,"_35":460,"_23":24,"_152":-5,"_283":461,"_26":462},{"_42":446,"_32":447,"_30":448,"_145":234,"_40":41,"_39":449,"_147":58,"_37":450,"_13":149,"_34":20,"_35":451,"_23":24,"_152":-5,"_283":452,"_26":453},{"_42":437,"_32":438,"_30":439,"_145":234,"_40":41,"_39":440,"_147":24,"_37":441,"_13":442,"_34":20,"_35":443,"_23":24,"_152":-5,"_283":444,"_26":445},{"_42":429,"_32":430,"_30":431,"_145":234,"_40":41,"_39":319,"_147":58,"_37":432,"_13":433,"_34":20,"_35":434,"_23":24,"_152":-5,"_283":435,"_26":436},{"_42":420,"_32":421,"_30":422,"_145":234,"_40":41,"_39":423,"_147":58,"_37":424,"_13":425,"_34":20,"_35":426,"_23":24,"_152":-5,"_283":427,"_26":428},{"_42":412,"_32":413,"_30":414,"_145":234,"_40":41,"_39":319,"_147":58,"_37":415,"_13":416,"_34":20,"_35":417,"_23":24,"_152":-5,"_283":418,"_26":419},{"_42":404,"_32":405,"_30":406,"_145":234,"_40":41,"_39":407,"_147":58,"_37":408,"_13":149,"_34":20,"_35":409,"_23":24,"_152":-5,"_283":410,"_26":411},{"_42":395,"_32":396,"_30":397,"_145":234,"_40":41,"_39":398,"_147":58,"_37":399,"_13":400,"_34":20,"_35":401,"_23":24,"_152":-5,"_283":402,"_26":403},{"_42":387,"_32":388,"_30":389,"_145":234,"_40":41,"_39":390,"_147":58,"_37":391,"_13":242,"_34":20,"_35":392,"_23":24,"_152":-5,"_283":393,"_26":394},{"_42":383,"_32":33,"_30":31,"_145":234,"_40":41,"_39":14,"_147":24,"_37":384,"_13":62,"_34":20,"_35":36,"_23":24,"_152":-5,"_283":385,"_26":386},{"_42":375,"_32":376,"_30":377,"_145":234,"_40":41,"_39":378,"_147":58,"_37":379,"_13":380,"_34":20,"_35":36,"_23":24,"_152":-5,"_283":381,"_26":382},{"_42":367,"_32":368,"_30":369,"_145":234,"_40":41,"_39":370,"_147":58,"_37":371,"_13":229,"_34":20,"_35":372,"_23":24,"_152":-5,"_283":373,"_26":374},{"_42":360,"_32":361,"_30":362,"_145":234,"_40":41,"_39":319,"_147":58,"_37":363,"_13":149,"_34":20,"_35":364,"_23":24,"_152":-5,"_283":365,"_26":366},{"_42":351,"_32":352,"_30":353,"_145":234,"_40":41,"_39":354,"_147":24,"_37":355,"_13":356,"_34":20,"_35":357,"_23":24,"_152":-5,"_283":358,"_26":359},{"_42":342,"_32":343,"_30":344,"_145":234,"_40":41,"_39":345,"_147":58,"_37":346,"_13":347,"_34":20,"_35":348,"_23":24,"_152":-5,"_283":349,"_26":350},{"_42":333,"_32":334,"_30":335,"_145":234,"_40":41,"_39":336,"_147":58,"_37":337,"_13":338,"_34":20,"_35":339,"_23":24,"_152":-5,"_283":340,"_26":341},{"_42":325,"_32":326,"_30":327,"_145":234,"_40":41,"_39":328,"_147":58,"_37":329,"_13":149,"_34":20,"_35":330,"_23":24,"_152":-5,"_283":331,"_26":332},{"_42":316,"_32":317,"_30":318,"_145":234,"_40":41,"_39":319,"_147":58,"_37":320,"_13":321,"_34":20,"_35":322,"_23":24,"_152":-5,"_283":323,"_26":324},{"_42":224,"_32":225,"_30":226,"_145":-5,"_40":41,"_39":227,"_147":24,"_37":313,"_13":229,"_34":150,"_35":230,"_23":24,"_152":-5,"_283":314,"_26":315},{"_42":217,"_32":218,"_30":219,"_145":-5,"_40":41,"_39":220,"_147":58,"_37":310,"_13":62,"_34":150,"_35":222,"_23":24,"_152":-5,"_283":311,"_26":312},{"_42":209,"_32":210,"_30":211,"_145":-5,"_40":41,"_39":212,"_147":58,"_37":307,"_13":214,"_34":150,"_35":215,"_23":24,"_152":-5,"_283":308,"_26":309},{"_42":201,"_32":202,"_30":203,"_145":-5,"_40":41,"_39":204,"_147":58,"_37":304,"_13":206,"_34":20,"_35":207,"_23":24,"_152":-5,"_283":305,"_26":306},{"_42":193,"_32":194,"_30":195,"_145":-5,"_40":41,"_39":196,"_147":58,"_37":301,"_13":198,"_34":150,"_35":199,"_23":24,"_152":-5,"_283":302,"_26":303},{"_42":185,"_32":186,"_30":187,"_145":-5,"_40":41,"_39":188,"_147":58,"_37":298,"_13":190,"_34":150,"_35":191,"_23":24,"_152":-5,"_283":299,"_26":300},{"_42":177,"_32":178,"_30":179,"_145":-5,"_40":41,"_39":180,"_147":24,"_37":295,"_13":182,"_34":150,"_35":183,"_23":24,"_152":-5,"_283":296,"_26":297},{"_42":169,"_32":170,"_30":56,"_145":-5,"_40":41,"_39":171,"_147":24,"_37":292,"_13":173,"_34":174,"_35":175,"_23":24,"_152":98,"_283":293,"_26":294},{"_42":161,"_32":162,"_30":163,"_145":-5,"_40":41,"_39":164,"_147":24,"_37":289,"_13":166,"_34":150,"_35":167,"_23":24,"_152":-5,"_283":290,"_26":291},{"_42":154,"_32":155,"_30":156,"_145":-5,"_40":41,"_39":157,"_147":58,"_37":286,"_13":62,"_34":150,"_35":159,"_23":24,"_152":-5,"_283":287,"_26":288},{"_42":142,"_32":143,"_30":144,"_145":-5,"_40":41,"_39":146,"_147":58,"_37":282,"_13":149,"_34":150,"_35":151,"_23":24,"_152":-5,"_283":284,"_26":285},["D",1772897660393],"children",[],[143],["D",1777759288465],[],[155],["D",1782848955587],[],[162],["D",1781346741270],[],[170],["D",1782848890794],[],[178],["D",1761591401040],[],[186],["D",1761591443489],[],[194],["D",1771009079911],[],[202],["D",1782848233329],[],[210],["D",1782848940289],[],[218],["D",1782844059093],[],[225],"834b8896-2f4a-11f1-a46b-42dde27851f2","x402.md","8fd15a24-1a3b-11f1-8274-42dde27851f2",869,["D",1775213250135],41,"---\ntitle: What if you never had to get an API key ever again?\ntags: ai, coding\ndate: 2026-03-08\n---\n\n# What if you never had to get an API key ever again?\n\nI care a lot about developer experience, so it kills my soul to watch someone\nget an API key.\n\nI'm the founder of [Val Town](https://val.town/), a developer platform, so I\nspend a lot of time watching people run into this.\n\nIn the age of vibe coding, getting API keys is particularly egregious. It takes\nyour agent 5 minutes to write the code, but takes you 30 minutes to get the key.\nMost of the time of making an app is getting API keys!\n\nThe worst part is that before you can try an API, you usually have no idea:\n\n1. How long sign-up will take\n2. Whether you'll need a credit card\n3. Whether the API is actually useful\n\nYou're at the mercy of the provider, and you just pray they care about developer\nexperience.\n\n### Workarounds\n\nThis problem has bothered me enough that we've tried our best to route around it\nat Val Town.\n\nThe easiest workaround is to push users towards free APIs that don't require\nkeys, like open-meteo for weather.\n\nFor premium APIs, we've created a series of hosted proxies that give users a\ntaste of the API without needing a key. We call this our \"Standard Library\", and\nit includes [sending emails](https://docs.val.town/reference/std/email/),\n[OpenAI](https://docs.val.town/reference/std/openai/),\n[Blob Storage](https://docs.val.town/reference/std/blob/), and\n[SQLite](https://docs.val.town/reference/std/sqlite/). We're also experimenting\nwith a proxy for [Clay](https://blog.val.town/clay) (for enrichments), and in\nthe past had a\n[Twitter proxy](https://www.val.town/x/stevekrouse/socialDataProxy).\n\nBut that puts Val Town in the unenviable position of rate-limiting, fighting\nabuse, etc, in an ad-hoc way. It's annoying and doesn't scale.\n\n### The Dream\n\nThe dream is:\n\n1. Users show up to any API with cash\n2. Send micropayments to try stuff out\n3. Minimal processing fees\n4. Open protocol > closed app store\n\nThink \"OpenRouter for everything. But as a protocol\".\n\nSurprise! It already exists.\n\n### x402\n\nIt's called [x402](https://www.x402.org/).\n\nAnd yes, unfortunately, this is the part where I tell you it has to do with\ncrypto.\n\nI regret to inform you that we may have found a real use-case for crypto.\n\nx402 is a new protocol created by Coinbase in May 2025 for making paid HTTP\nrequests. It's built on top of the pre-existing 402 HTTP status code,\n`Payment Required`.\n\nAny HTTP API server can now optionally support x402 by returning\n`402 Payment Required`, explaining the price, and what crypto wallet to send\npayment to.\n\nThe client can then make a payment on the blockchain, and retry the request with\nthe proof of payment. The server verifies the payment, and release the premium\nAPI response.\n\n```mermaid\nsequenceDiagram\n participant Client\n participant Server\n participant Blockchain\n Client->>Server: HTTP Request\n Server->>Client: 402 Payment Required\n Client->>Blockchain: Payment\n Blockchain->>Client: Payment Confirmed\n Client->>Server: HTTP Request (with proof of payment)\n Server->>Client: HTTP Response\n```\n\nThe code is elegant. You add your wallet's private key to your env vars, and\nthen wrap your fetch with the x402 client. The SDK takes care of the payment\nflow for you.\n\n```ts\nconst signer = privateKeyToAccount(process.env.PRIVATE_KEY);\nconst client = new x402Client();\nregisterExactEvmScheme(client, { signer });\n\nconst paidFetch = wrapFetchWithPayment(fetch, client);\nconst res = await paidFetch(\n \"https://x402.browserbase.com/browser/session/create\",\n {\n method: \"POST\",\n headers: { \"Content-Type\": \"application/json\" },\n body: JSON.stringify({ duration: 5 }),\n },\n);\n```\n\nYou can think of it as attaching a wallet to `fetch` for it to autopay for\npremium requests.\n\nThis is the dream. We can make a paid a API request without ever going to that\nAPI's website, signing up, getting an API key, attaching a credit card, etc. In\nthe above example, we're just showing up to Browserbase with a penny, and they\nare giving us a websocket URL to a 5-minute browser session.\n\n[You can try out a running demo of this flow here.](https://www.val.town/x/stevekrouse/x402-playground)\n\n### Accepting payments\n\nOnce I was able to make paid requests, I wanted to see how hard it was to accept\npayments.\n\nIn theory, because it's a protocol, not an app store, it should just be a couple\nlines of code to hang my shingle, and accept payments into the same wallet I\ncreated to make payments.\n\nSo I made the world's most expensive `Math.random()`. If you show up with a\ntenth of a penny, I will give you a random number.\n\n```ts\nconst seller = privateKeyToAccount(process.env.PRIVATE_KEY);\nconst payTo = seller.address;\n\napp.use(\n paymentMiddleware(\n {\n \"GET /api/random\": {\n accepts: {\n scheme: \"exact\",\n price: \"$0.001\",\n network: \"eip155:8453\", // Base\n payTo,\n },\n description: \"A premium random number\",\n mimeType: \"application/json\",\n },\n },\n resourceServer,\n ),\n);\n\napp.get(\"/api/random\", (c) => c.json({ random: Math.random() }));\n```\n\n[You can try out a running demo of this flow here.](https://www.val.town/x/stevekrouse/x402-playground)\n\n### Getting a wallet\n\nIf you want the full walkthrough, including the wallet setup and all the rough\nedges, I made a [video version](https://www.youtube.com/watch?v=lqr4FIZdR9s).\n\nNow, to be clear, the getting a wallet and adding USDC is still not great. It\ntook me an annoying 20 minutes the first time, and even the second time, when I\nknew what I was doing, it still took 10 minutes.\n\nBut the shape of this feels right. And, importantly, this is a one-time cost and\nit should unlock _all the world's APIs_ from this single private key. It's like\nI got a debit card for APIs and now I can walk into any API store and swipe it\nimmediately without first pre-registering.\n\n### The brave new world\n\nI think the end state looks something like this:\n\n1. You sign up to Val Town (or Replit, Lovable, Vercel, etc)\n2. The platform creates a wallet for your app, or lets you import one\n3. The platform gives you a few free pennies to play with (or you add your own\n via credit card or whatever)\n4. You vibe code an app in minutes; it uses who-cares-how-many APIs\n5. It works. The app is done in minutes. You never got any API keys.\n\nThe beauty is you don't even have to know about which APIs are being used\nanymore, _because you don't have to click around in a web browser to get API\nkeys_. Your agent can just start making paid requests to prototype with any API\nthat supports x402.\n\nWe need more sellers on x402. I'm honestly confused why there isn't more of a\ngold rush here. Anyone can take any premium API – OpenAI, Anthropic, etc – and\ncreate a paid proxy with x402 and start earning money. There's an OpenRouter\nshaped opportunity here to be the reputable proxy on top of all the world's APIs\n– until the API providers themselves catch up and put themselves on x402, like\nBrowserbase did.\n\nWe also need the crypto wallet onramp to get much easier. The experience of\ngetting crypto, transferring it to a new wallet, etc, is still pretty bad. The\nbetter the onramp, the more buyers and sellers we'll get, and the faster this\nwill grow.\n\nHere at Val Town, we're continuing to watch x402 developments with great\ninterest, with an eye towards integrating these wallets into our platform\nnatively. The dream of never again getting an API key is close at hand.\n\n_Edit: [dt](https://thyresson.io/) made a good point that this infra could also\nhelp end-users pay for API costs directly. For example, I could make an app that\nuses Browserbase (and a dozen other x402 premium APIs), but then _not_ provide\nmy own developer app wallet. Instead my app could ask ask my _users_ to provide\ntheir own wallet to pay for their usage of these services directly. I could even\nimagine a \"Login with Coinbase\" flow to provide your wallet credentials\nsecurely. This way your app could go viral without breaking the bank or having to deal with payments, rate-limits, etc._",[],[233,317],"3abe2d20-55df-11f1-a5da-ee650bb23af1","survey.md","cd1b7e86-1a3c-11f1-be9a-42dde27851f2",991,["D",1779455267667],"---\ntitle: Survey the space\ntags: advice, creativity\ndate: 2025-03-24\n---\n\n# Survey the space\n\n![a person surveying a space](https://stevekrouse-blob_admin.web.val.run/api/public/survey.jpeg)\n*\"a person surveying a space\" by [Fal Demo App](https://www.val.town/v/stevekrouse/falDemoApp)*\n\n> 🤖 **This is an AI draft.** I asked Claude to take a stab at this essay\n> based on the [stub](#the-original-stub) below and on the technique I've\n> been teaching it to use on Twitter. Steve will rewrite it eventually.\n> Until then, read it as a transcript of my thinking, not as my prose.\n\nIn [Find a model](/model), I argued that the single highest-leverage move\nwhen you're making anything — a tweet, an essay, a pitch deck, a UI, a song —\nis to find an exemplar to keep on your desk while you work.\n\nThe natural follow-up question: *how do you find a good model?*\n\nThe honest answer is that it's the hard part. Most people skip it, pick the\nfirst thing that comes to mind, and end up modeling their work on whatever\ntheir brain happened to surface in the first 30 seconds. That's usually not\nthe best model in the world. It's just the most available one.\n\nThe technique for actually finding a good model is what I call **surveying\nthe space**. It's pretty much what it sounds like — but it's worth being\ndeliberate about, because the quality of your model bounds the quality of\nyour output.\n\n## The basic move\n\nSurveying the space means doing real research before you start making the\nthing. Not \"I'll google it if I get stuck\" research — actual,\nlook-up-front, take-notes research. Here's the loop:\n\n- **Describe precisely what you're doing.** Not \"a tweet about AI\" but\n \"a tweet that shares a surprising chart from privileged data, aimed at\n the Claude Code team, that might earn a retweet from them.\" The more\n specific your description, the better your search queries will be.\n- **Think of the key words.** Brainstorm the vocabulary that the people\n who already made the thing you want to make would use. If you're\n writing about onboarding, the word \"activation\" might find a different\n corpus than the word \"onboarding.\" Cast wide.\n- **Think about where this crowd hangs out.** A great model for a\n Hacker News post lives on Hacker News. A great model for a viral\n developer tweet lives on Twitter. A great model for an essay that gets\n read for a decade lives on someone's blog, not in a Substack feed.\n Different ponds, different fish.\n- **Search everywhere.** Google, Perplexity, HN search, Twitter advanced\n search, Deep Research, GitHub, your own bookmarks. Each surface has\n different blind spots — Google buries Twitter, Twitter buries blog\n posts, Perplexity buries niche communities. Triangulate.\n\nThat's the abstract version. Let me walk through a concrete one.\n\n## Worked example: finding a model for a tweet\n\nThe flavor of surveying I've done the most is for tweets. I've gotten\nopinionated enough about it that I've been teaching Claude to do it for me.\n\nSay I want to tweet a chart from internal Val Town data, and I want a\nspecific audience — say, the Claude Code team — to retweet it. The naive\napproach is to write the tweet and hope. The slightly-less-naive approach\nis to search Twitter for \"Claude Code chart\" and copy whatever's trending.\n\nThe real move is to **find a tweet shaped like the one I want to write\nthat already earned a retweet from the people I want to amplify me**. If\nthey retweeted that shape once, they're probably willing to retweet that\nshape again. The shape is pre-validated. That's the model.\n\nHere's how I survey the space for that, broken into three parallel\nhunts I'd run before drafting a single word:\n\n1. **Topic search.** Broad Twitter advanced-search queries on the genre\n — \"AI coding chart\", \"developer survey\", whatever vocabulary the\n community uses. This finds tweets that performed well on the topic\n *in general*. Good baseline, but not pre-validated for my specific\n amplifier.\n2. **Author timelines.** A curated list of authors I respect in this\n space — the people whose tweets I'd be flattered to be mistaken for —\n and pull their recent tweets with `from:author` queries. These give\n me shape candidates from people whose taste I already trust.\n3. **Amplifier RTs.** This is the highest-signal one and the one I'd\n never do by hand because it's annoying. Pull the *timeline* of the\n amplifier account (the Claude Code team's account, in this example),\n extract their recent retweets and quote-tweets, and filter for ones\n that match the shape I want. A tweet they've already retweeted is a\n tweet they're already willing to retweet.\n\nAfter the three hunts come back, I look at maybe 30–50 candidate tweets,\npick 3–5 that feel closest to the shape I want, and *then* I start\ndrafting. The first draft is almost always better than what I would've\nwritten from scratch, because I'm not staring at a blank page — I'm\nstaring at a target.\n\n## How an agent can do this for you\n\nMost of the steps above are mechanical. \"Pull these timelines, filter\nfor retweets, group by shape\" is not a creative act. It's a search-and-\nfilter job, and search-and-filter is exactly what agents are good at.\n\nIf you give an agent access to the Twitter API — or, much easier,\n[twitterapi.io](https://twitterapi.io) (an unofficial wrapper that costs\nfractions of a cent per call and doesn't require Twitter's developer\nprogram) — the entire survey becomes a single prompt. The agent can run\nthe three hunts in parallel, dedupe results, score candidates by\nengagement, and hand you a ranked list of models in a few minutes.\n\nI've been doing this in Claude Code with a small skill. The skill is\nmaybe 30 lines of instructions plus a script that calls twitterapi.io's\nadvanced-search and user-timeline endpoints. The instructions are the\nimportant part. Here's the gist of what I tell the agent:\n\n> **Skill: survey-the-space-for-a-tweet**\n>\n> Before drafting a tweet, run three parallel hunts via twitterapi.io:\n>\n> 1. **Topic search.** Run 3–5 Twitter advanced-search queries on the\n> genre. Brainstorm vocabulary the target community actually uses;\n> don't just paraphrase my words.\n> 2. **Author timelines.** From a curated list of authors the user\n> respects, pull recent tweets with `from:` searches. Look for\n> shape matches, not topic matches.\n> 3. **Amplifier RTs.** Pull the timeline of the target amplifier\n> account. Extract their retweets and quote-tweets from the last 90\n> days. Filter for ones matching the requested shape (chart-share,\n> surprising-finding, comparative-claim, etc.).\n>\n> Surface 3–5 candidate models to the user with links and short\n> rationales. Do **not** synthesize them into one \"average\" model — the\n> point of a model is that it's a specific concrete artifact. Let the\n> user pick.\n>\n> After the user picks, draft in two rounds: first round, 5 deliberately\n> *different* shapes; second round, 5 variants *within* the chosen\n> shape.\n\nThe reason the skill works is that it encodes the structure of the\nsurvey, not the content. The content — vocabulary, authors, amplifiers —\ngets refilled per task. That's the part I'd want a human collaborator to\ndo anyway.\n\nIf you want to try this without writing the skill yourself: get a\ntwitterapi.io key, stash it in an env var, and ask Claude Code (or any\nagent with shell access) to \"survey Twitter for tweets shaped like X\nthat have been retweeted by Y.\" It'll figure out the rest. Each hunt\nruns in a sub-agent in parallel, which means the whole survey finishes\nin roughly the time of the slowest single API call.\n\n## Why this generalizes\n\nThe Twitter version is the one I've automated, because Twitter has an\nAPI and tweets have shapes. But the same loop works for almost anything\ncreative:\n\n- **Essays:** search blog archives, RSS feeds, \"best of\" lists, and\n HN top posts on the topic. Look for essays with the structure you want,\n not just the topic.\n- **Pitch decks:** Pitch.com has a public gallery; so does\n bestpitchdeck.com. Search by industry *and* by round size.\n- **Product copy:** swipe-file sites like marketingexamples.com and\n greatlandingpagecopy.com exist. Use them.\n- **Code architecture:** search GitHub for production projects in the\n same shape as yours — not tutorials, not toy examples, *production*\n projects. Read their `README`, their `package.json`, their top-level\n directory layout.\n\nThe unifying move in all of these is: figure out where the people who\nalready made the thing hang out, and look at what they made.\n\n## The deeper point\n\nSurveying the space feels like procrastination because you're not making\nthe thing yet. You're just looking. It feels like the \"real work\" is\nproducing the artifact.\n\nThat feeling is wrong. The real work is choosing what to produce, and\nyou can't choose well without seeing what's already out there. Half an\nhour of surveying saves five hours of producing something nobody wanted.\n\nIf you have an agent that can survey for you, the cost of surveying\ndrops to roughly zero, which means there's no excuse not to do it\nanymore.\n\n## The original stub\n\nFor posterity, here's the bulleted stub this draft was built from:\n\n- *Describe precisely what you're doing*\n- *Think of key words*\n- *Think of words or phrases you'd be likely to use*\n- *Think about where you'd post your thing. Think about where this\n crowd hangs out*\n- *Google, perplexity, hn search, twitter, deep research*\n",[],[233,326],"f0ef9f4c-1a3a-11f1-9195-42dde27851f2","request-for-school.md","1395c7c4-f948-11f0-9da0-42dde27851f2",718,["D",1772897587771],21,"---\ntitle: Request for School\ntags: education, parenting\ndate: 2026-02-19\n---\n\n# Request for School\n\n_Location: Brooklyn, NY_\n\nEver since I went to a magical after-school program\n([IMACS](https://www.imacs.org/)) that helped me go from hating to falling in\nlove with math, I've been extremely passionate about education. To learn about\neducation, I've read dozens of books, visited schools, did a lot of tutoring and\nteaching, and eventually started my own after-school computer science program\nhere in NY, [The Coding Space](https://www.thecodingspace.com/).\n\nMy biggest educational influences are Seymour Papert, The Sudbury Valley School,\nBlake Boles, John Holt, and John Taylor Gatto.\n\nI believe that humans learn best at their own pace, voluntarily,\nself-directedly, so they have a chance to develop an intimate, positive\nemotional relationship with the material.\n\nI believe that a lot of what traditional schools do is actively harmful to\nchildren: large classrooms, curriculum, lectures, tests, homework, etc. These\none-sized-fits-all tools are useful for school systems to manage mass-education\nat scale, but they are woefully worse than more humane and tailored approaches.\n\nFor example,\n[Bloom's Two Sigma Problem](https://en.wikipedia.org/wiki/Bloom%27s_2_sigma_problem)\nshows that literally all students can get A's in all subjects if instead of\ngetting lectured at, they get a personalized tutor. Erik Hoel writes at length\nabout how\n[we used to produce Einsteins through what he calls Aristocratic Tutoring](https://www.theintrinsicperspective.com/p/why-we-stopped-making-einsteins).\nEspecially with modern technology, such as the internet, computers, google,\nwikipedia, not even including recent advances in AI, there are definitely\ncost-effective ways to get every child the attention they deserve. For starters,\nthe flipped classroom model works wonders; that's what I experienced 20 years\nago at IMACS. Secondly, if you remove the overhead of paying for extensive\nadministration of a school, then you can much more affordably scale 1:4 teacher\nto student ratios.\n\nThis is why homeschooling is on the rise. It's growing at an insane rate right\nnow!\n\nYet, it feels depressing to keep a child alone at home all day, learning with\njust mom, and on the computer by themselves.\n\nIt feels like there should be another option that's a vibrant, social place for\nchildren to go that's a culturally rich environment to learn, be creative, make\nfriends, go on adventures, do projects, play make-believe, have fun, learn\nlessons, have space for academic rigor and play, silliness and serious\nproductive fun, time to run around outside and focus time inside, time to read\nwhatever you want, time to go on a walk with friends.\n\nGiven how much I care about this, I fear that I might have to create this\n\"school\" myself: provide this vision, recruit the leaders, teachers, other\nparents and students, to help get it off the ground. If this sounds like your\nlife's work, please reach out!\n\nHowever, I'd be much happier to find that a school close enough to my dream,\nclose enough to where we want to live, and just get to send our kids there,\nwithout having to start it ourselves! **Would you know of such a school?**\n\nIs a Montessori school close enough to what I'm looking for? What about Waldorf?\nAre there any other good pre-existing options in my area?\n\nMy fiancé Emily and I are planning to raise a family in Brooklyn, NY. We\ncurrently live in Prospect Heights. We plan to have 2-3 kids.",[],[233,334],"41db72a8-9e5b-45db-bf5a-1a6b70466160","readme.md","d8c137a4-b19f-11f0-847d-0224a6c84d84",1003,["D",1782300160792],27,"---\ntitle: Steve Krouse's README\ntags: personal, management\ndate: 2025-10-25\n---\n\n# Steve Krouse's README\n\n👋 Hi, I'm Steve. This README explains a bit about me, the standards I hold\nmyself to and how to work best with me. Hopefully, reading this document will\nsave you some time learning my quirks the old-fashioned way. If you find me not\nliving up to my own standards, please let me know, so I can make it right.\n\n## A bit about me\n\n- I was born in NYC and raised in South Florida\n- I'm a big audiobook and podcast listener\n- I dropped out of Penn after 3 semesters\n- I'm Bret Victor and Alan Kay fanboy\n- I'm passionate about programming and education\n- I'm a Crystal Palace supporter\n\n## Values\n\nThese are some of the standards I hold myself to, and traits I value in others.\n\n1. **Integrity** – honesty, accountability, doing the right thing. Doing what\n you'll say you'll do. Not lying. Not taking advantage of people. Being\n responsible. Integrity is the cornerstone of trust, and trust lets us all go\n further, faster.\n\n2. **Kindness** – treating people with respect, patience, good-will. Wanting the\n best for others. Taking pleasure in helping people. Taking others'\n perspectives and emotion into account, and treating them with care and\n empathy. Working _with_ people instead of manipulating them.\n\n3. **Resourcefulness** – leveraging all resources and clever thinking to\n identify and solve problems. Not getting stuck, spiraling, or getting\n discouraged. Asking for help when you need it – other people are some of the\n best resources! – but not bothering people until you've done your homework\n first.\n\n## Async communication\n\nI love async communication.\n\nI regularly check:\n\n- email\n- text\n- facebook messenger\n- whatsapp\n- twitter DMs\n- bluesky\n- linkedin\n\nI try to at least acknowledge receipt of all non-spam messages within a day or\ntwo. Always feel free to bump a message to urge me to reply more quickly.\n\nI'm a big fan of sending and receiving audio or video messages when it makes\nsense more than text. For example, for giving feedback on an essay or a website,\ndemoing a bug, or teaching how to do something.\n\nI may message you at odd hours, but that doesn't mean I expect a reply right\naway. That's what's so great about async communication: the sender can send\nmessages at the right time for them, and the receiver can reply at the right\ntime for them, and those can be two very different times.\n\nWhen there's a tense, fraught, complex, or important conversation happening on\nDMs, I try to quickly bounce to synchronous conversation, like phone call. A lot\ncan be lost in textual communication that can be easily cleared up with the tone\nof a human voice.\n\n## Giving feedback\n\nI love giving feedback. My favorite way is async, stream-of-conscious style,\nwherein I document my reactions to the work, sometimes in comments, sometimes in\na video recording. It's how I review my own work. There's minimal sugar coating.\nI may highlight parts and simply say \"wordy\" or \"nonsense\" or \"delete\" or\n\"repetitive\" or \"word choice\" or \"did an AI write this?\".\n\nPlease interpret these as **reactions**. I am not claiming to know what's right.\nI am not demanding you make any change.\n\nTake my feedback with a grain of salt. I am just one person. I may not be your\nintended audience. I can be wrong. Maybe I misread something. My reaction is a\nsingle data point. I'm letting you know what goes inside the head of one person\nwhen they read your work. My reactions do not need any reply. They do not\nrequire any action on your part. If you think my reaction is irrelevant, ignore\nit! Let my reactions wash over you, and then do what you think is best.\n\n## Receiving feedback\n\nI love to receive feedback, and really appreciate it when it's given in any\nform. I see it as my responsibility to make it as easy and painless for you to\ngive me feedback.\n\nI absolutely love friction logs. They're one of the best gifts you can give me.\n[Here's a great guide on how to write one](https://blog.sbensu.com/posts/friction-logs/),\nbut don't feel like yours has to be so structured. If easier, just turn on your\ncamera, screen recorder and mic, and talk me through your feedback.\n\n## Rough edges\n\nI have a couple of rough edges that I am aware of. While I'm working on them and\nhope you don't have to experience them, I find it helpful for folks to know what\nthey are just in case.\n\nIf I'm making you feel bad, it's likely because I'm making you feel flustered or\ndumb. I hope I never do this to you. If I do, know that I am not doing it on\npurpose, and am sorry about it. Likely it means that I'm feeling stupid or\noverwhelmed, and am trying to protect myself by projecting the stupid onto you.\nIf you're brave and self-aware enough to say, \"You're making me feel stupid\",\nthose are magic words to fix the situation. It will help me adjust my own\nattitude, de-escalate, and apologize.\n\nAlso know that my first reaction to an idea is not the final word on it, so\ndon't try to read into it too much. If you think I'm wrong, that's great! It\nmeans that there's something for one or both of us to learn. I admire people who\ncan self-advocate for themselves and their beliefs. If I am being dismissive of\nyou or your ideas, please call me out on it, so I can give them the focus they\ndeserve.\n\n## Tools\n\nThis one is a bit random, but I want everyone I work with to use the following\ntools, or equivalents. Otherwise, it's just annoying to watch you use your\ncomputer 😝. They are quick to install. Then you just have to pick keyboard\nshortcuts and learn them. It's a very worthwhile investment. You won't know how\nyou lived without them.\n\n1. [Clipboard manager](https://maccy.app/) - if you think cmd+c is useful, wait\n till you can cmd+c multiple things\n\n2. [Window manager](https://rectangleapp.com/) - for looking at things\n side-by-side.\n\n3. [Screen Recording](https://cleanshot.com/) - for async communication. You can\n skip a lot of meetings with a good screen recording.\n\n## Parting ways\n\nI want to build a work culture where we celebrate people for:\n\n1. **Long tenures**. Folks should feel like there's no artificial ceiling to\n learning, growing, and compensation on my team.\n2. **Parting on good terms**. There's not always room for everyone's dreams at\n every company, and we want to celebrate our people chasing their dreams,\n wherever that takes them.\n\nIf you're thinking of quitting on me, I'd first ask that you let me know that\nyou're unhappy or thinking of a move as early as possible so that I can do all\nthat I possibly can to fix things to make you happy to stay. Even if it seems\nimpossible, please let me know, because you might be wrong. If it doesn't seem\nlike I understand the urgency of your request, let me know that too.\n\nTypically, when an employee wants to leave an employer, they keep that knowledge\nsecret until they are ready to hand-in their two-weeks notice. This is very\nsensible, considering that the employer may decide after hearing this to\nimmediately fire the employee or otherwise treat the employee badly.\n\nI find this pattern sub-optimal and would like to propose an alternative to\nanyone I manage: trust me like your mentor that has your best interests at\nheart. If you are unhappy working with me or plan to leave, let me know as soon\nas possible, so that I can address it or prepare. In return for this trust, I\ncan promise that I will treat you right. I won't fire you or be upset. On the\ncontrary, I'll be incredibly grateful, and even more helpful you in your next\nsteps.\n\nI have the receipts. I told this to an employee who later took me up on it. She\neven asked me to even be her reference for her future employer. I gave a glowing\nreference, she got the job, and we had a very pleasant final weeks together. If\nyou'd like to speak to her as a reference, let me know! I hope to treat you so\nwell in this situation that you too will be an employee reference for me in the\nfuture.",[],[233,343],"6589ed92-aea6-430d-9908-c086d806569b","quotes.md","430e1e56-051a-11f1-b3bd-42dde27851f2",1006,["D",1782397815453],32,"---\ntitle: Favorite Quotes\ntags: philosophy\ndate: 2025-03-01\n---\n\nA list of my favorite quotes.\n\n## Learning\n\n> Only rarely does some exceptional event lead people to reorganize their\n> intellectual self-image in such a way as to open up new perspectives on what\n> is learnable\n>\n> - Seymour Papert, Mindstorms\n\n## Precision\n\n> If you’re thinking without writing, you only think you’re thinking.\n>\n> – Leslie Lamport\n\n> The purpose of abstraction is not to be vague, but to create a new semantic\n> level in which one can be absolutely precise.\n>\n> — Edsger Dijkstra\n\n> Instead of regarding the obligation to use formal symbols as a burden, we\n> should regard the convenience of using them as a privilege: thanks to them,\n> school children can learn to do what in earlier days only genius could\n> achieve.\n>\n> When all is said and told, the \"naturalness\" with which we use our native\n> tongues boils down to the ease with which we can use them for making\n> statements the nonsense of which is not obvious.\n>\n> – Edsger W.Dijkstra,\n> [On the foolishness of \"natural language programming\"](https://www.cs.utexas.edu/~EWD/transcriptions/EWD06xx/EWD667.html)\n\n> In order to make machines significantly easier to use, it has been proposed\n> (to try) to design machines that we could instruct in our native tongues. This\n> would, admittedly, make the machines much more complicated, but, it was\n> argued, by letting the machine carry a larger share of the burden, life would\n> become easier for us. It sounds sensible provided you blame the obligation to\n> use a formal symbolism as the source of your difficulties. But is the argument\n> valid? I doubt.\n>\n> – Edsger W.Dijkstra,\n> [On the foolishness of \"natural language programming\"](https://www.cs.utexas.edu/~EWD/transcriptions/EWD06xx/EWD667.html),\n> [h/t cratermoon](https://news.ycombinator.com/item?id=47480530)\n\n> Everything is vague to a degree you do not realize till you have tried to make\n> it precise.\n>\n> – Bertrand Russel\n\n> The quantity of meaning compressed into a small space by algebraic signs, is\n> another circumstance that facilitates the reasonings we are accustomed to\n> carry on by their aid.\"\n>\n> – Charles Babbage, quoted in Iverson's Turing Award Lecture, from\n> [Succinctness is Power by Paul Graham](https://paulgraham.com/power.html)\n\n> There are two ways of constructing a software design: One way is to make it so\n> simple that there are obviously no deficiencies, and the other way is to make\n> it so complicated that there are no obvious deficiencies.\n>\n> –\n> [Tony Hoare](https://cs.stanford.edu/people/eroberts/courses/soco/projects/2008-09/tony-hoare/index.html),\n> [h/t Paul Buchheit](https://news.ycombinator.com/item?id=47325225)\n\n# Augmentation as prosthetic\n\nFrom Plato's Phaedrus,\n[by way of fs](https://fs.blog/an-old-argument-against-writing/):\n\n> Here, O king, is a branch of learning that will make the people of Egypt wiser\n> and improve their memories. My discovery provides a recipe for memory and\n> wisdom. But the king answered and said ‘O man full of arts, the god-man Toth,\n> to one it is given to create the things of art, and to another to judge what\n> measure of harm and of profit they have for those that shall employ them.’\n>\n> And so it is that you by reason of your tender regard for the writing that is\n> your offspring have declared the very opposite of its true effect. If men\n> learn this, it will implant forgetfulness in their souls. They will cease to\n> exercise memory because they rely on that which is written, calling things to\n> remembrance no longer from within themselves, but by means of external marks.\n>\n> What you have discovered is a recipe not for memory, but for reminder. And it\n> is no true wisdom that you offer your disciples, but only the semblance of\n> wisdom, for by telling them of many things without teaching them you will make\n> them seem to know much while for the most part they know nothing. And as men\n> filled not with wisdom but with the conceit of wisdom they will be a burden to\n> their fellows.\n\nFrom Bret Victor & Alan Kay in\n[Re: using AI to augment human intelligence](https://dynamicland.org/archive/2017/Is_this_the_civilization_we_really_want):\n\n> I would argue that a tool that augments one's intelligence must leave that\n> person with a better understanding of whatever domain they're working in. If\n> the tool develops an \"understanding\", but does not induce that understanding\n> into the user, I don't think it can be called an IA tool.\n>\n> One of Alan's favorites analogies is the prosthesis on a healthy limb. Here's\n> one phrasing of it (from here):\n>\n> Put a prosthetic on a healthy limb and it withers. Using the logic of current\n> day education, we could say that since students are going to be drivers as\n> adults, at age two we should put them in a little motorized vehicle and they\n> will just stay there and learn how to be much better drivers. Now, we would\n> think that was pretty horrible. But what if we gave the same person a bike?\n> We're not going to feel so badly [because] the bike allows that person to go\n> flat out with his body and it amplifies that. [The bike is] one of the great\n> force amplifiers of all time because it doesn't detract from us--it takes\n> everything we've got and amplifies it. Most computers today are sold like\n> cars, where as many things as possible are done for you. You don't have to\n> understand how it works and, in fact, you don't have to understand how to\n> think because the most popular stuff is prepackaged solutions for this and\n> that. When you put a person into a car, their muscles wither. You put a person\n> into an information car, and their thinking ability withers. I wouldn't put a\n> person within 15 yards of a computer unless I was absolutely sure that it was\n> a kind of a bike for them.\n>\n> The most important question for an IA tool is: does it grow the person? Or\n> does it create a dependency which ultimately weakens them?\n>\n> Turkle's \"Simulation and its Discontents\" has some powerful observations about\n> students and practitioners who were weakened in this way by computational\n> tools. They could produce work faster (for some definition of \"work\"), but\n> knew less and understood less.\n>\n> (And, among other things, would make egregious \"what were you even thinking\"\n> mistakes. Because the tool had prevented them from learning to think.)\n\n# Misc\n\n> Creators need an immediate connection to what they're creating\n>\n> – Bret Victor\n\n---\n\n> malloc for the internet\n>\n> – Jeff Bezos's original spec for s3\n\n---\n\n> A meeting is a refuge from ‘the dreariness of labor and the loneliness of\n> thought.’”\n>\n> — Bernard Baruch, “A Theory on Meetings” (1970), quoted in The Design of\n> Design by Frederick P. Brooks Jr.",[],[233,352],"834b5e20-2f4a-11f1-a46b-42dde27851f2","proposal.md","cb8235ce-1a3c-11f1-becb-42dde27851f2",["D",1775213250135],"---\ntitle: I proposed to my girlfriend in Quordle\ntags: coding\ndate: 2025-02-17\n---\n\n# I proposed to my girlfriend in Quordle\n\n![Emilrdle screenshot](https://stevekrouse-blob_admin.web.val.run/api/public/emilrdle1.png)\n\nEmily loves puzzles. Every day she plays Wordle, the New York Times Mini, Connections, Spelling Bee, LinkedIn Tango and Queens, and Quordle, in that order.\n\nI thought it'd be fun to propose in one of her favorite games. It's appropriate, after all, because Wordle itself was created for love. But which game would work?\n\nAt the end of the day you have to say \"Will you marry me?\" so that ruled out many options. How can you fit that into Wordle or Connections? I tried my hand at a crossword, but damn those are hard!\n\nEventually I realized the question would fit inside Quordle, but with extra letters. For context, [Quordle](https://www.merriam-webster.com/games/quordle/#/) is four Wordles that you're playing at the same time. I'd deal with the extra letters by having them fall away at the end, leaving the big question. The first set of words I came up with were \"Willy\", \"Youth\", \"Marry\", and \"Meaty\", but I eventually changed \"Meaty\" to \"Meter\" because Emily *loves* when words end with \"er\".\n\nEnter AI. I work at [Val Town](https://val.town), and we make a website for programmers to build apps. [Townie](https://val.town/townie) is our AI-assistant that writes Val Town apps for you. I started with a very simple prompt and a screenshot of Quordle.\n\nOn a cold December night at Betaworks in Chelsea, I cozied up and worked with Townie for 90 minutes. It was one of David Bieber's wonderful hack nights, so the atmosphere was perfect. By the end of the night, it was done. I remember it being a lot of fun.\n\nLooking back at the chat history, I see that there are [91 versions](https://www.val.town/v/stevekrouse/emilrdle/versions), which means I was averaging about one edit per minute. It takes Townie, powered by Claude 3.5 Sonnet, about 30-60 seconds to rewrite a couple hundred lines of code, so that would account for the speed. I think most of those edits were through Townie, but I also made some manual edits to the code for smaller changes.\n\nThe basic mechanics of Quordle were more complicated than I'd realized. For example, the keys on the on-screen keyboard need to be colored appropriately, green, yellow, gray, and black, representing whether that letter is in the right spot, wrong spot, not in the word, or not played yet, respectively, and in four quadrants, for the four different Wordle boards on-screen. Also each of the attempts on each Wordle board needs to be colored appropriately, with subtle rules governing various end states.\n\nBut after I got the basic mechanics working, the game felt flat somehow. I pulled up the original and noticed all sorts of tiny micro interactions that gave the game life. When you press on a letter on the keyboard, it moves up and turns another color. When a letter shows up on the screen, it animates in. When you get a word correct, it animates in in a flipboard style.\n\nI also wanted to copy how Quordle highlights invalid words in red, but for that I needed a list of valid words. It was suprisingly easy to find the list that Quordle itself was using. I popped open the Chrome Network tab, navigated to Quordle, and hit Cmd+F to get the Search across everything. There I just put a random word that I knew to be on the list, \"youth\", and voila. I found it interesting to see the lists, broken up into a `wordBank`, `blacklist`, and `allowed`. I went with the `allowed` list.\n\nI later realized that the letters-falling-away animation at the end happened quite quickly, so I needed some way to slow it down, and get her attention. I could imagine her looking away right when she gets the last word, and she'd miss it. So I had Townie add a modal when she won, asking her if she's ready for the next level.\n\n\n

\n \n

\n\nThe morning I decided to do it, we first started with her normal word games. Then I announced that I made her a word game. She was excited. She enjoyed the Quordle, but was a bit confused why all the words were so random. The only one that had any significance was \"meter\", because it ended in \"er\", but even that significance was weak. Safe to say she had no clue what was coming. She got to the win screen and was ready for the next level. The letters fell away and she turned to me and said, \"What?\"\n\nThe rest happened in a blur, but I did get down on one knee and pull out a ring. She said yes.\n\n

\n\nYou can play [Emilrdle here](https://stevekrouse-emilrdle.web.val.run/). The next level is reserved for Emily.\n",[],[233,361],"5d5f0e12-e57a-43d5-845c-76fc0916edd9","precision.md","f71a5698-2564-11f1-b6b7-42dde27851f2",1009,["D",1782837209377],"---\ntitle: Reports of code's death are greatly exaggerated\ntags: ai, coding, writing\ndate: 2026-03-21\n---\n\n# Reports of code's death are greatly exaggerated\n\nMarch, 21 2026\n\n[A sufficiently detailed spec is code](https://haskellforall.com/2026/03/a-sufficiently-detailed-spec-is-code)\nbegins with this lovely comic:\n\n![A four-panel CommitStrip comic. Panel 1: A man holding a coffee mug says to a colleague typing on a laptop, \"Some day we won't even need coders any more. We'll be able to just write the specification and the program will write itself.\" Panel 2: The colleague replies, \"Oh wow, you're right! We'll be able to write a comprehensive and precise spec and bam, we won't need programmers any more!\" The first man sips his coffee and says, \"Exactly!\" Panel 3: The first man asks, \"And do you know the industry term for a project specification that is comprehensive and precise enough to generate a program?\" The colleague answers, \"Uh... no....\" Panel 4: The colleague says, \"Code,\" and then, \"It's called code.\"](https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/f7e9e621-0fad-4c3a-a65b-d16832c02b00/public)\n\nThere is a profound tension here: english specifications _intuitively feel_\nprecise until you learn better from bitter experience. (It's all in that facial\nexpression of the last frame.)\n\n> \"Everything is vague to a degree you do not realize till you have tried to\n> make it precise.\"\n>\n> – Bertrand Russel\n\nProgramming, like writing, is an activity, where one iteratively sharpens what\nthey're doing as they do it. (You wouldn't believe how many drafts I've written\nof this essay.)\n\nAI helps you with this, because it – increasingly instantly and well – turns\nEnglish into running code. You can then react to it – \"move the button there;\nmake it bluer\" – to get incrementally more precise about what you want.\n\nThis is why \"vibe coding\" is such a perfect phraseology: you stay operating at\nthe level of your English-level vibes while reacting to the AI-created artifacts\nthat help you sharpen your thinking.\n\nBut, vibe coding gives the illusion that your vibes are precise abstractions.\nThey will feel this way right up until they\n[leak](https://www.joelonsoftware.com/2002/11/11/the-law-of-leaky-abstractions/),\nwhich will happen when you add enough features or get enough scale. Unexpected\nbehaviors (bugs) that\n[emerge from lower levels of abstraction](https://worrydream.com/LadderOfAbstraction/)\nthat you don't understand will sneak up on you and wreck your whole day.\n\nThis was Dan Shipper's experience when his\n[vibe-coded text-editor app went viral, and then went down](https://every.to/chain-of-thought/when-your-vibe-coded-app-goes-viral-and-then-goes-down).\nAs it turns out, \"live collaboration is just insanely hard.\"\n\n

\n\n\"Live collaboration\" _intuitively feels_ like a perfectly precise specification.\nWe've all used Google Docs, Notion, etc so it _feels_ precisely spec'd. It's\nincredibly hard a priori to see what this is not the case.\n\nThe only reason that I personally know otherwise is that I tried to add a\ncollaborative text editor to a product I was working on 10 years ago, and it was\nan unexpected nightmare of complexity.\n\nWhat was hard about it? I don't remember! That's part of the problem! Complexity\ncan be incredibly boring, unpleasant to think about, and hard to remember all\nthe details and edge cases. For example, the classic flowchart of how Slack\ndecides when to send you a notification:\n\n![image.png](https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/4370b0ce-5657-4edc-9077-25751bbdd500/public)\n\nBut, this isn't the end of the story either. We are blessed with an extremely\npowerful tool to master complexity.\n\n## Abstraction\n\nThere is a fundamental limit in the human brain. We can only think of 7 (plus or\nminus 2) things at a time. So the **only way** to think about more than 7 things\nis to _compress_ multiple things into a single thing. Happily, we can do this\nrecursively, indefinitely, which is why humans can master unlimited complexity.\nThat compression step is called _abstraction_.\n\n> The purpose of abstraction is not to be vague, but to create a new semantic\n> level in which one can be absolutely precise.\n>\n> — Edsger Dijkstra\n\nFor example, Sophie Alpert used clever abstraction to\n[refactor the Slack diagram to this much simpler one](https://sophiebits.com/2024/10/30/everyone-is-wrong-about-that-slack-flowchart):\n\n

\n\nThis is the best part of programming: coming up with increasingly good\nabstractions to help us master complexities. My favorite examples of this are\nfunctional programming concepts, like functional reactive programming, which\n[I wrote a wonderful essay on](https://futureofcoding.github.io/futureofcoding.org/essays/dctp).\n\nSo yes, collaborative text editors are fundamentally complex, but that just\nmeans that we're continually in search of better abstractions to help us master\ncomplexities, like ReactJS or TailwindCSS did in their respective domains.\n\n## AGI\n\nBut let's play this out 1, 2, 5, 10, 100 years. AI is getting\nbetter/faster/cheaper at incredible rates, but regardless of _when_, unless you\nbelieve in magic, it's only a matter of time until we reach the point at which\nmachine intelligence is indistinguishable from human intelligence. We call that\npoint AGI.\n\nIt may seem like an AGI world is a vibe world. If anyone can afford 100\nKarpathy-level geniuses for $1000 / month, why ever trouble yourself with any\ntroublesome details? Just have your army of Karpathys handle them for you.\n\nThis is such a joke to me. This is _clearly_ only something you'd think in the\nabstract, before this technology arrived.\n\nIf you told me that I had access to that level of intelligence, there is zero\npart of me that is going to use it to ship _more_ slop. Are you freaking\nkidding?? Of course not.\n\nI think we're confused because we (incorrectly) think that code is only for the\nsoftware it produces. It's only partly about that. The code itself is also a\ncentrally important artifact. When done right, it's poetry. And I'm not just\nsaying this because I have Stockholmn Syndrome or a vested interest in it – like\na horse jockey might in the face of cars being invented.\n\nI think this is a lot clearer if you make an analogy to writing. Isn't it\nfucking telling that nobody is talking about \"vibe writing\"?\n\nWe're not confused with writing because there's nothing mystical about\nsyntactically correct sentences in the same way there is about running code.\nNobody is out there claiming that ChatGPT is putting the great novelists or\njournalists out of jobs. We all know that's nonsense.\n\nUntil we get AGI. Then, by definition, machines will write amazing non-slop and\nit'll be glorious.\n\nThe same exact situation is true for coding. AI produces (increasingly less)\nshitty code. We all know this. We all work around this limitation. We use AI\nin spite of the bad code.\n\nAs Simon Willison says,\n[AI should help us produce better code](https://simonwillison.net/guides/agentic-engineering-patterns/better-code/).\nAnd when we have AGI this will be easy.\n\nWhen we have AGI, the _very first_ things we will use it on will be our hardest\nabstraction problems. We will use it to help us make better abstractions so that\nwe can better understand and master complexity.\n\nYou might think the need for good code goes away as AIs get smarter, but that's\nlike using ChatGPT to write more slop. When we get AGI, we will use them to make\nbetter abstractions, better collaborative text editor libraries, etc.\n\nFor example, my favorite success story with Opus 4.6 was that it helped\nme with my dream full-stack react framework for Val Town. It one-shot solved\n[my list of unsolved problems](https://www.val.town/x/stevekrouse/reactRouter7Example/branch/main/version/164/code/react-router-questions.md)\nthat I had with getting React Router 7 to work full-stack in Val Town. The\nresult is my nascent [vtrr](https://www.val.town/x/stevekrouse/vtrr) framework.\nI'm particularly proud of this 50 line full-stack react app demo in\n[a single file](https://www.val.town/x/stevekrouse/vtrr/code/demos/hello-world.tsx):\n\n![Screenshot 2026-03-21 at 18.48.14@2x.png](https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/448cbc3e-a196-4d2f-7f96-6b0dcf822200/public)\n\nIf you know of any other snippet of code that can master all that complexity as\nbeautifully, I'd love to see it.\n\n## Reports of code's death are greatly exaggerated\n\nIt seems like 99% of society has agreed that code is dead. Just yesterday I was\nlistening to podcaster Sam Harris of all people confidently talking about how everyone agrees\ncoding is dead, and that nobody should learn to code anymore.\n\nThis is so sad. It's the same as thinking storytelling is dead at the invention\nof the printing press. No you dummies, code is just getting started. AI is going\nto be such a boon for coding.\n\nI have so much more to say on this topic, but this essay is already 3x longer\nthan I wanted it to be. I'll stop here and leave you with some of my favorite\nquotes on formalism.\n\n> Instead of regarding the obligation to use formal symbols as a burden, we\n> should regard the convenience of using them as a privilege: thanks to them,\n> school children can learn to do what in earlier days only genius could\n> achieve.\n>\n> When all is said and told, the \"naturalness\" with which we use our native\n> tongues boils down to the ease with which we can use them for making\n> statements the nonsense of which is not obvious.\n>\n> – Edsger W.Dijkstra,\n> [On the foolishness of \"natural language programming\"](https://www.cs.utexas.edu/~EWD/transcriptions/EWD06xx/EWD667.html)\n\n> There are two ways of constructing a software design: One way is to make it so\n> simple that there are obviously no deficiencies, and the other way is to make\n> it so complicated that there are no obvious deficiencies.\n>\n> –\n> [Tony Hoare](https://cs.stanford.edu/people/eroberts/courses/soco/projects/2008-09/tony-hoare/index.html),\n> [h/t Paul Buchheit](https://news.ycombinator.com/item?id=47325225)\n\n> The quantity of meaning compressed into a small space by algebraic signs, is\n> another circumstance that facilitates the reasonings we are accustomed to\n> carry on by their aid.\"\n>\n> – Charles Babbage, quoted in Iverson's Turing Award Lecture, quoted in\n> [Succinctness is Power by Paul Graham](https://paulgraham.com/power.html)",[],[233,368],"b894e91d-c0d1-4828-a207-52c81e182d81","phoenix.md","8698ea9a-74ad-11f1-ab90-1607ee4eb77e",1084,["D",1782922887234],53,[],[233,376],"1a4993d8-686d-47f7-ab2e-275c14bf691f",["D",1783370560154],[],[233,33],"0cd5bc1c-f4a4-11f0-a39f-42dde27851f2","nvc.html","a05348a0-f2e8-11f0-a6db-42dde27851f2",523,["D",1768764587512],"\n \n NVC\n \n \n \n

\n \n What are you feeling?\n

\n \n

Peaceful

calm

content

satisfied

secure

centered

relieved

confident

comfortable

composed

fulfilled

Affectionate

warm

tender

appreciative

passionate

compassionate

loving

touched

Happy

glad

pleased

excited

joyful

delighted

amused

encouraged

grateful

hopeful

thrilled

moved

Playful

energetic

expansive

adventurous

mischievous

silly

light

hearted

enthusiastic

lively

optimistic

Interested

inspired

curious

surprised

stimulated

amazed

absorbed

fascinated

engaged

eager

enchanted

intrigued

motivated

Rested

alert

alive

invigorated

refreshed

rejuvenated

relaxed

renewed

Sad

lonely

hopeless

hurt

sorrowful

grief

despair

discouraged

disappointed

helpless

Confused

hesitant

indifferent

suspicious

puzzled

troubled

bewildered

baffled

conflicted

torn

uncertain

skeptical

Scared

terrified

startled

shocked

nervous

anxious

worried

desperate

afraid

cautious

insecure

Uncomfortable

uneasy

embarrassed

fragile

vulnerable

ashamed

distressed

disturbed

bored

restless

jealous

impatient

exasperated

stressed

envious

tense

concerned

tired

weary

exhausted

overwhelmed

Angry

irritable

annoyed

agitated

frustrated

disgusted

furious

enraged

mad

angry

bitter

\n \n \n What are you needing?\n \n \n

Subsistence

air

food

health

movement

safety

rest/sleep

shelter

touch

water

Security

consistency

efficiency

predictability

order

structure

protection

stability

accountability

trust

reassurance

Freedom

autonomy

choice

ease

independence

responsibility

space

spontaneity

empowerment

self-expression

Well-being

health

healing

rest

privacy

self-care

nurturing

Relaxation

humor

joy

play

fun

recreation

pleasure

rejuvenation

Connection

affection

appreciation

attention

closeness

companionship

harmony

intimacy

love

support

tenderness

warmth

To Matter

acceptance

care

compassion

consideration

empathy

kindness

respect

to be heard

to be seen

understanding

Community

belonging

communication

cooperation

equality

inclusion

mutuality

participation

partnership

shared reality

Awareness

authenticity

competence

dignity

growth

honesty

integrity

clarity

discovery

learning

Meaning

aliveness

challenge

consciousness

contribution

creativity

effectiveness

exploration

integration

purpose

Transcendence

beauty

balance

celebration

communion

faith

flow

hope

inspiration

mourning

peace

presence

wholeness

\n \n \n \n",[],[233,388],"8d0f4acd-389d-465f-80a7-ee0359cfbb66","model.md","f264fbf0-1a3b-11f1-a11c-42dde27851f2",1008,["D",1782832965276],3,"---\ntitle: Find a model\ntags: advice\ndate: 2025-03-24\n---\n\n# Find a model\n\n![a woman painting a landscape](https://stevekrouse-blob_admin.web.val.run/api/public/model.jpeg)\n_\"a woman painting a landscape\" by\n[Fal Demo App](https://www.val.town/v/stevekrouse/falDemoApp)_\n\nOne of the most common pieces of advice I give to mentees is to \"find a model.\"\nIt's a stupid-simple technique, but can transform how you work. It will let you\ndo things faster, better, and with more confidence. It's a mystery that\neverybody doesn't use it all the time.\n\nThe core of the technique is to find a model artifact to keep an eye on while\nyou work on your artifact, which could be a coding project, a song, a painting,\nan email, a social media post, an essay like this one, etc.\n\nThe are three important benefits to this technique.\n\n## A north star\n\nA model gives you a north star; something to shoot to; something you can\nvisually compare your work to, and see what's missing; a yardstick by which you\ncan measure how \"done\" your artifact is. Without a model, how do you even know\nwhat kind of thing you're making? How can you tell its shape? A model is like a\nhyper-specific template that lets you stay focused on the end goal.\n\nThere are many related principles:\n[Backward Chaining](https://en.wikipedia.org/wiki/Backward_chaining),\n[Top-down design](https://en.wikipedia.org/wiki/Bottom-up_and_top-down_design),\n[The XY Problem](https://en.wikipedia.org/wiki/XY_problem), and Amazon's\nprinciple to \"Write the Press Release First\". These all get at the idea that you\ndon't want to lose sight of the end goal. Without a model, you risk getting lost\nin the details. Losing the forest for the trees. A model helps you keep the\nwhole forest in sight, so you actually create what you set out to.\n\nHowever, it's important to remember that \"Plans are useless, but planning is\nindispensable.\" In other words, you shouldn't feel compelled to stick to your\nmodel so closely. Use the parts that you want, and forget about the rest. A\nmodel is high up on [The Ladder of Abstraction](), but that doesn't mean you\ncan't change it. If you realize that you chose the wrong model or that your\ngoals have changed, excellent. Time to find a better model.\n\nThe fact that you have a model is what allows you to realize when you have the\nwrong model. Imagine you're a ship captain navigating by a star that you thought\nwas the north star, but then later realized was just Venus. The time you wasted\nnavigating by Venus would've been equally wasted if you were navigating without\nreference to any star. At least when you're navigating by Venus one of your\nfellow sailors can come over your shoulder and say, \"Hey captain, I think that's\nVenus.\" If you're navigating by nothing, then you're truly lost at sea.\n\n## Un-stuck-ing yourself\n\nAnother reason to have model is that it's very helpful when you get stuck. Don't\nknow how to do something? See how your model does it.\n\nFor this essay I felt stuck on the first line. This essay's model is\n[Do things that don't scale](https://www.paulgraham.com/ds.html), because it's\nan essay that explains a cognitive technique. Its first sentence isn't the\nprettiest, but it gets the idea across and it's a hook: \"One of the most common\ntypes of advice we give at Y Combinator is to do things that don't scale.\"\n\nNow go back and read my first line. Is it deritive? Yes. Would literally anyone\nhave noticed if I didn't draw your attention to it? No. There's something\nprofoundly true about the line oft-attributed to Picasso, \"Good artists borrow.\nGreat artists steal.\" I didn't truly believe this quote until I read\n[The Elements of Eloquence](https://en.wikipedia.org/wiki/The_Elements_of_Eloquence),\nwherein he shows examples of Shakespeare lifting whole sentences out of other\nbooks and plays, only changing a couple words.\n\nIn school, they call this behavior cheating, which is probably why this\ntechnique isn't practiced more often. I think it's an important lesson to\nunlearn. Why reinvent the wheel? Save time so you can focus on your unique\ncontribution. Just be sure to give attribution and credit where it's due.\n\n## Maybe someone already did it, so you don't have to\n\nHow many times have you built something only later to realize someone else\nalready built it? Chances are that if an idea occurs to you, likely it has\noccured to someone like you. So many people don't even take a second to google\naround before embarking on a project. Instead of wasting your time reinventing\nthe wheel, see if someone already invented it. Then stop. Just use their thing.\n\n## How to find a model\n\nOne reason \"finding a model\" isn't more common is that it can be quite difficult\nto find a good model. I call that technique [\"surveying the space\"](./survey).\n\n## Further reading\n\n- [Stealing is a Skill](https://ben-mini.com/2026/stealing-is-a-skill)",[],[233,396],"053162c4-f509-458f-95c5-1066dfb6437e","legibility.md","8d8127ce-6ccd-11f1-a5fd-1607ee4eb77e",1000,["D",1781976666983],"---\ntitle: A love letter to legibility\ntags: legibility, writing\n---\n\n_June 20, 2026 Note: This post is in-progress. What you see below is a very\nrough first draft_\n\n# A love letter to legibility\n\nLegibility is the ease with which something can be understood. It's an\nincredibly useful mental model that will make you smarter, help you get what you\nwant, and is a lens to better understand the world.\n\nLegibility means giving the world affordances for how they can interact with you\nand fit you into their lives – for jobs, romance (have you heard of Date Me\nDocs?), friends, collaborators, vendors, etc. You can think of it as your\n\"interface\", like your own personal API. Or as Henrik says, \"A blog post is a\nvery long and complex search query to find fascinating people and make them\nroute interesting stuff to your inbox\".\n\nBy default, you're invisible to the world. At the other end of the spectrum are\ntruly famous people. I'm talking about this secret third thing where you put a\nlot of stuff about yourself and what you're interested in, working on, thinking\nabout, etc, online so that you are a visible person.\n\nWhen I started the Future of Coding podcast & community and blogged super\npublicly, I thought about creating a lighthouse to attract my people and\ncollaborators. I walked into an empty room and started talking at the walls and\nyou'll never guess what happened next. People came into the room! And they're\nstill there! (Even though I left!)\n\nLegibility is how I found some of my best friends, investors, employees and\nsolutions to my problems. Unfortunately Twitter is not what it was, but find a\nplace where you can post online – a link is all that matters – and do as much of\nit as you can. Quantity over quality, excluding slop, of course.\n\nA well-phrased problem is a problem half solved. In framing it, you sometimes\nsolve it. Now that it's written down you can email it to friends, tweet it, ask\nAI. I had incredible back pain and was complaining about it to a friend who had\na super kooky solution that worked out.\n\nHowever, don't let illegibility stop you from doing things, especially in the\nearliest days. When I started my independent researcher / podcaster phase,\neveryone in my life thought I was crazy, and it was hard to explain to folks\nwhat I was doing. I just kept talking about it online in whatever way I could,\nand it got more legible over time. Henrik does a great job explaining the\nimportance of illegibility in Looking for Alice.\n\nTake date me docs. Making yourself legible in a dating way, most importantly to\nyourself – getting clearer about what you want vs not, almost like a rubric, so\nyou don't get fooled by love feelings and just go with the person you're with.\nImportant to do before you fall in love. A hire me doc is similarly important,\nso you know what you want, so you don't take the first job offered like so many\ndo.\n\nIt's important if you have these to keep yourself honest to what you want; and\neven more powerful if you can empower your friends and family to keep you honest\nwith what you want. I remember my brother telling me what he wants in a job and\nthen a month later going for a job that was the opposite – but it wasn't written\ndown, and there wasn't an understanding that I was supposed to help him stick\nwith what he wanted originally.\n\nThere's a nuance that the person who first writes such a doc is the most\nignorant version of you, and that future versions of you will know more – but\nthat's for updating your doc! You should treat it as a living document that\nyou're constantly referencing for yourself, constantly showing people, getting\nfeedback on. This is why it's so important that it be short and fun and legible\n– so people can read it quickly, you can keep it in your head, perfect and\nagonize over every word. If it's a static document you made once, it doesn't do\nthe trick.\n\nThis relates all the way back to when you were a kid and your math teacher\ninsisted you show your work on the math test. Showing your work is incredibly\npowerful. One of the most powerful cognitive techniques. On a meta level,\nlearning to show your work is way more important than any math you were\nlearning. When I used to teach my brother Will debugging, I would tell him he\nneeds to externalize his thinking. Once something is externalized it can be\ndebugged, addressed, reacted to. When your thinking is trapped in your own head,\nit's not inspectable, it's not legible.\n\nI'm reading \"What is Intelligence?\" right now, and it talks about how just like\nhumans, LLMs don't know what they're thinking, because they too are inscrutable\ndistributed systems, until they say what they're thinking (by predicting the\nnext token). And this is why chain of thought is so incredibly valuable for both\nLLMs and humans: it allows our thinking to be legible to ourselves! It's a way\nfor us to know what we're thinking. It allows our whole brain to get on the same\npage and think the same chain of thought together. I'm doing this right now – I\nhave so much more clarity about what I'm thinking now because I'm thinking it\nout loud on paper.\n\nThis style of writing, this stream of consciousness thing I'm doing right now,\nis an example of making yourself legible to yourself. Very very powerful. So a\nhuge part of why legibility is important is that it's one of the most powerful\ncognitive techniques for thinking, just in solo mode.\n\nBut then we layer in multiplayer mode and the benefits explode! Date me docs\nmake you legible to the dating marketplace – which of course includes other\npeople you might date, but sometimes even more importantly, it includes people\nwho can introduce you to those people, or give you advice on how to find those\npeople! The same goes for resumes or hire-me docs. This is very clear in India\nwhere they have dating resumes. There's also a practice on Twitter that I love,\nabout people bearing their soul about the job they want or the person they want\nto hire. It seems like such an incredibly effective way to get what you want.\nWhen something is legible it tells the world at large how to interact with it;\nhow they can fit it into their life, what they need.\n\nThat brings me to interfaces. Something is legible if you know how to interact\nwith it; what its interfaces are; what the inputs and outputs are. What it\ncosts. How to talk to it. How to get help when you get stuck. How to navigate\nit. What it even is.\n\nSo this brings me to APIs. That's a very precise kind of legibility. Or UIs,\nanother precise kind of legibility. Or a website for a product. Or a map of a\ncity. Or a map of a subway system. Or a directory when you walk into a building.\nOr the front of the building – where is the door. A legible subway station would\nmake it easy to know how to get on the right train, and hard to accidentally get\non the wrong one. When you arrive on a floor of a hotel and are looking for room\n321, is it immediately obvious if you turn left or right? Can you take it all in\nin a glance?\n\nThis brings me to Tufte, and the visual display of quantitative information.\nSuper relevant. Also Bret Victor of course, and representations that make things\nunderstandable. His mission is to build media for understanding. You could call\nthat \"legible media.\" Learnable Programming is a great essay on legible\nprogramming. A learnable or a legible interface have a lot in common – maybe\nthey're the same even.\n\nAnd then at Val Town I've been working a lot to make us as a company legible:\nwho works here, how much funding we raised, our pricing and business model.\nProduct legibility: what it is, who it's for, what problem it solves. This is\ntypically called positioning, and it's all about legibility. Supabase is super\nlegible and that's part of its success: it was an open-source Firebase\nalternative on Postgres. Super duper legible.\n\nWhich brings me to analogies. I love analogies and metaphors, and all the books\nabout them from Hofstadter to Lakoff – those really help with legibility. Which\nbrings us to Seymour Papert, who talks about how the ease with which you can\nlearn a new thing depends upon the models already in your head. So legibility is\nbottlenecked by the audience. It's easier to make a map for someone who can read\nmaps!\n\nIt's a powerful mental model like \"growth mindset\" or \"grit\" or \"debugging.\" I'm\ngiving you a word to describe an important and overlooked thing, which – now\nthat you have a word for it – you can do better.\n\nI have this thesis that the bottleneck for human progress (always, but\nespecially in the age of LLMs) is human understanding. We will soon have a\ncountry of geniuses in a datacenter, but we will be bottlenecked by\nunderstanding what they're doing and directing them appropriately. I used to\nthink this was a problem of getting the supergenius AIs to slow down and explain\nthemselves to us poor dumb humans, getting their genius into our thin skulls.\nBut I no longer think this. It's actually externalizing their thinking to\nthemselves that makes LLMs as smart as they are, and will make them smarter.\nRecently I've been forcing LLMs to make things legible to me. I've put their\n\"effort\" level to low, and asked them to think out loud with me, react to my\nthoughts and to their own thoughts, getting their thoughts out in public (not\nprivate like LLMs typically do) so that we can both get smarter as a single\nsystem.\n\nPeople have this scary view of inscrutable intelligent systems that are\nintelligent in ways we could never understand. That's nonsense! The very nature\nof intelligence is that it's intelligible – aka legible. Richard Feynman has a\ngreat quote about how you only understand something if you can explain it to a 5\nyear old. The path to making LLMs smarter is to make them better at explaining\nthings to themselves and others. This is rubber duck debugging. Or why you think\nbetter in conversation. (Human intelligence may have evolved for persuasion and\ngroup think, not thinking in isolation.)\n\nEvery powerful mental model also has downsides. The world is really messy and\ncomplicated. It's our responsibility and opportunity to make sense out of it. So\nto answer the downside: it's to not confuse the map with the territory. To not\nmake a legible version coercive. To let people be messy, or have their own kinds\nof legibility. Henrik Karlsson's \"Looking for Alice\" essay is great – about how\nto avoid legibility early on in romance, and this was also helpful for me. The\nsame is true in the earliest days of a startup. If you are too focused on\nlegibility, you can't feel into something and let it bloom out of confusing\ncomplexity. This is related to when you abstract too early in software. It's\nbetter to start concrete, and then once you've done something enough, you can\nabstract and make legible. It's common for an abstraction to make something less\nlegible, which just means it's a bad abstraction. Bret Victor's \"Up and Down the\nLadder of Abstraction\" is key. Just because you can think in one way doesn't\nmean you should do it all the time.\n\nMaybe everything is legibility. We exist in a world spread out in time and\nspace, and at every level of resolution, we have entities which are looking out\nfor other entities, and legibility is what allows these entities to decide which\nway to move, whether to interact, how to interact – everything from cells\ninteracting, to predator/prey, to collaboration, to how different brain regions\nmake themselves legible to other brain regions. Making yourself legible to the\nworld is a form of hooking yourself up to the human superorganism, the\nsuperbrain. Doing the same thing on the internet really hooks you up to the\ndigital nervous system of the human superorganism. There's something beautiful\nabout that.",[],[233,405],"834ac208-2f4a-11f1-a46b-42dde27851f2","ir.md","915cfc4c-1741-11f1-8f7c-42dde27851f2",["D",1775213250135],18,"---\ntitle: Personal Investor Relations\ntags: startups, legibility, writing\ndate: 2026-03-03\n---\n\n# Personal Investor Relations\n\nFor most companies, investor relations and communication are unpleasant hygiene, like a dental cleaning, something to get through as fast and painlessly as possible, often delegated to automations, ops, finance, or IR. The way some founders do it, it's a beautiful art form all unto itself: clear, plain-spoken prose, minimal corporate-ese doublespeak, humor, often self-deprecating, lessons learned, and plans for the future.\n\nI am thinking first and foremost of [Warren Buffett](https://www.berkshirehathaway.com/letters/letters.html), but this would also include [Jeff Bezos & Amazon](https://www.aboutamazon.com/about-us/shareholder-letters). For example:\n\n> “Mistakes — yes, we make them at Berkshire… Don’t beat yourself up over past mistakes — learn at least a little from them and move on.”\n> — [Warren Buffett](https://www.berkshirehathaway.com/letters/letters.html)\n\n> “I believe we are the best place in the world to fail (we have plenty of practice!), and failure and invention are inseparable twins.”\n> — [Jeff Bezos](https://www.aboutamazon.com/news/company-news/amazons-original-1997-letter-to-shareholders)\n\nMore recently, I've discovered that [Patrick and Jon of Stripe](https://stripe.com/annual-updates/2025) and [Immad of Mercury](https://mercury.com/blog/annual-letter-2025) also publish lovely letters. What they share is a willingness to explain how they think, what they got wrong, and what they plan to do next, in their own voices.\n\nTo me, these kinds of investor letters fit into the broader legible universe that includes [Date Me Docs](https://dateme.directory/) (which I am proudly the world's leading advocate for), beautifully written job postings, and founder-written blogs, like [Paul Graham's](https://www.paulgraham.com/). In all of these, the common thread is creators speaking plainly and directly to the world.\n\nThe are big benefits to taking investor relations and communications seriously. First, it helps you build trust:\n\n1. **Build trust with capital markets.** Over time, they learn how you think, act, and allocate capital. That earns you a longer leash to make long-term investments, more grace when you have missteps, and often higher multiples than competitors who stay opaque. Something that surprises many people is that public markets actually include plenty of extremely long-term oriented investors. Buffett talks about this a lot, and I have many investor friends who think the same way.\n\n2. **Build trust with labor markets.** Employees, who are always implicit underwriters of a business and often shareholders themselves, see that you are trustworthy stewards of their careers, livelihoods, and reputations.\n\n3. **Build trust with customers.** They can see more behind the curtain how you are trying to serve them, not trick or exploit them.\n\n4. **Build trust with the public.** You explain why your company exists, what purpose it serves, and how you try to be a good community member. In this anti-business and anti-billionaire climate, some scorn is earned, but some of it comes from a lack of transparency that leads people to assume the worst. I’ve never heard anyone say anything bad about Warren Buffett, and there’s some public relations genius there worth studying.\n\nThere are also a couple benefits that don’t quite fit under “trust,” but matter just as much:\n\n1. **Attract the partners you want and repel the ones you don’t.** Buffett talks a lot about the importance of aligned partners for the long term. The last thing I want is an investor looking to make a quick buck or take the company in a different direction, or a customer or employee who wants us to be something we don’t want to be.\n\n2. **Better internal decision making and alignment.** Spending time on investor communications makes oneself and one’s company more legible internally. (Sometimes I joke that the Date Me Directory is a dating app whose best feature is that it \"tricks\" people to be more thoughtful about who they are and who they want to be in a relationship with. Internal legibility, aka clarity, is worth the all by itself.)\n\nBut from an efficiency perspective, publishing thoughtful investor letters is a worthwhile use of time simply because you should be doing this reflective thinking anyway. And if you are going to spend the time reflecting, you might as well publish it and kill many birds with one stone: clarity for yourself, and legibility to everyone else. Often the only thing stopping you is a vague fear or making things more public than you need to. In my calculations, the benefits way exceed the costs for me, especially for the long-term, positive-sum games I intend to play for my career. Increased public legiblity helps others fit you into their stories where appropriate, increasing your “[luck surface area](https://www.codusoperandi.com/posts/increasing-your-luck-surface-area).”\n\nThis is why I plan to start publishing Val Town's Monthly Investor Updates. My very next Val Town investor update, which I’m writing now, is intended to be public, so hopefully within a few days folks can read it.",[],[233,413],"19887a3c-466e-11f1-be1c-42b51c65c3df","intelligence.md","f77b43da-10d9-11f1-ac03-42dde27851f2",921,["D",1777757411372],58,"---\ntitle: Intelligence Buying Intelligence\ntags: ai, startups\ndate: 2026-02-01\n---\n\n_May 2, 2026 Note: This post was written with too much AI assistance. Re-reading\nit now, it still feels like there are important ideas here, but it also feels\nvery slop-y. Apologies in advance! Hopefully I'll find time to write a proper\nhuman-typed version._\n\n# Intelligence Buying Intelligence\n\n_Feb 1, 2026_\n\n[Nnamdi Iregbulem recently wrote](https://whoisnnamdi.substack.com/p/tokens-arent-fungible)\nabout how users stick with expensive and slow closed models (ie Opus 4.6)\ndespite cheaper OSS alternatives.\n\nWhy aren't we using cheaper models?\n\nSonnet 4.6 is right there, faster, almost 2x cheaper, and about as good as Opus\n4.5, which we were all obsessed with five minutes ago – not to mention OSS\nmodels which are 10x cheaper. It's crazy that I'm always reaching for the\nslowest and most expensive model!\n\nFor me, there's extra cognitive dissonance, because not only do I care (like\neveryone else) about saving money, but I care especially about being in a flow\nstate. Waiting longer for a maybe-smarter response is the opposite of that. Back\nwhen Sonnet 3.5 was the best model, I couldn't wait to get my hands on something\nas smart as it but 10x faster and cheaper. Now we have that. Yet I never reach\nfor it.\n\nNow, I daydream of Opus 4.6 at Taalas's 17k tokens-per-second (Opus 4.6 is ~70\ntps, so 250x slower), yet I'm realizing that by the time we get there – say in a\nyear or two – will we even want to use it vs Opus 6.6?\n\nHere's my best explanation for why we keep reaching for the most expensive and\nslow models, even when cheaper and faster ones are available. There are four\nforces at play that make intelligence a weird thing to buy:\n\n1. The returns to intelligence are **nonlinear**\n2. Intelligence is **hard to measure**.\n3. Intelligence is **competitive**.\n4. Intelligence is a **commodity**. (You can always turn incremental capital\n into incremental intelligence.)\n\n## 1. The returns to intelligence are nonlinear\n\nThere's a new question tripping up most LLMs right now. It's the new \"how many\nr's in strawberry\":\n\n> The car wash is 50 meters away. Do you walk or drive?\n\nSonnet 4.6 and ChatGPT 5.2 tell you to walk.\n\nOpus 4.6 is the only model I know of that correctly tells you to drive (\"because\nyou'll need the car\"). Once you've seen a model make a mistake like that, it's\nreally hard to look it in the eyes and trust it with anything important.\n\nThis is because we know deep down that returns to intelligence are spiky and\ndiscontinuous. A tiny increase in intelligence can unlock a perspective that\nsimply doesn't exist at the level below it. The losses from insufficient\nintelligence are basically unbounded: someone who is fundamentally confused in\none seemingly small way can end up in the catastrophically wrong place. And the\nconverse is true too. Intelligence builds on itself. One correct insight early\non compounds into a totally different trajectory.\n\nConsider Warren Buffett as the poster child for this. He has a few key insights\nthat are basically invisible to most people, but those insights compound into a\ncompletely different trajectory. The difference between Buffett and the average\ninvestor may be a 0.001% difference in intelligence, but that can lead to\nmassively different outcomes.\n\nIntelligence isn't about exerting more effort. It's about seeing what's right in\nfront of you more clearly: finding the shortcut, the hidden door, the structural\nsimplification that was invisible a moment ago.\n\nIt's also about course-correcting as fast as possible when you realize you're on\nthe wrong path. Feedback loops dominate. Learning compounds.\n\nIntelligence might actually be unique in its nonlinearity. In a\n[David Deutsch](https://en.wikipedia.org/wiki/David_Deutsch) sense, it's the\nonly process that creates new explanatory knowledge, the creation of which is\ninherently unpredictable.\n\n## 2. Intelligence is hard to measure\n\nAnd here's what makes this extra tricky: we don't have good evals.\n\nThis is true for human evals too: SATs only goes up to 1600. If you have a room\nfull of 1600s, you'd have future Nobel laureates sitting next to future\ndeadbeats. And many Nobel laureates and future self-made billionaires and\ngeneration-defining-artists don't even get 1600s! Our tests literally don't\nmeasure the intelligence that matters most.\n\nLLMs keep exceeding our benchmarks, so we keep making harder ones. We're finally\ntaking teaching intelligence and measuring it seriously. I wish we were\nsimilarly ambitious about human intelligence: \"Oh damn, 40% of kids are getting\n1600s these days, gotta make SATs harder.\"\n\nWithout great evals, we default to vibes, reputation, brand loyalty, and price\nsignals. Which, honestly, is rational when you can't measure what matters.\n\nThe best eval I heard about recently is\n[Demis Hassabis's \"Einstein test\"](https://x.com/rohanpaul_ai/status/2025274123560362347):\ntrain an LLM on all human knowledge up to 1911, then see if it can independently\ndiscover general relativity. We need more evals like that.\n\n## 3. Intelligence is competitive\n\nThere's a third reason to always reach for the most expensive intelligence: in\nmany domains, intelligence is a race.\n\nI used to think that once we get models as smart as my CTO, Tom MacWright, we\ncan basically stop there and then work on compressing that intelligence down to\n8b parameters and making it faster and cheaper. Give me a country of Tom's in a\ndatacenter.\n\nBut in some domains, even if a model (or model + harness) is 0.001% smarter,\nthat could be the difference between winning and losing. In a competitive\ndomain, you don't just need to be good enough. You need to be better than\neveryone else.\n\nMy dad likes to tell the story of two hikers who see a bear. One hiker starts\nrunning, and the other says \"you can't outrun a bear!\" The first hiker says:\n\n> \"I don't have to outrun the bear, I just have to outrun you.\"\n\nIf someone could give you a consistent edge in the stock market, even a small\none, that edge could be worth a trillion dollars. But only if nobody else has\nit. The moment everyone has the same insight, it's priced-in and worthless. For\nthe stock market, you don't just need intelligence – you need _more_\nintelligence than everyone else.\n\nI'm not sure to what extent this is true in startups – classic YC advice is to\nignore competition. But in the attention economy, you definitely need to be\nbetter than everyone else. If your content is 0.01% better than the next best\nthing, that could be the difference between going viral and being ignored.\n\n## 4. You can always spend incremental dollars to get incremental intelligence\n\nRight now the most expensive models aren't actually that expensive. A few bucks\nper conversation. But that's changing fast.\n\nWe keep finding ways to spend more money for more intelligence: longer chains of\nthought, parallel reasoning, models that loop and self-correct for hours instead\nof seconds. Each of these multiplies the compute bill – on purpose.\n\nThere's a great example of this called the\n[Ralph Wiggum loop](https://ghuntley.com/ralph/), coined by Geoffrey Huntley.\nThe idea is simple: keep running an AI agent in a loop, feeding failures back\nin, until it succeeds. You're literally turning compute dollars into results\nthrough brute-force persistence. It's not elegant, but it works – and it means\nthat if the stakes are high enough, you can always throw more money at a\nproblem.\n\nWe're not just going to make intelligence cheaper. We're going to make it _more\nexpensive on purpose_ – because we'll keep finding ways to spend more compute on\nharder problems. The ceiling on what you can pay for intelligence is going to go\nway up. And if the returns are nonlinear, people will pay it. Gladly.\n\n## Nobody ever went bankrupt spending too much on good counsel\n\nA very-well-paid consiglieri once told me:\n\n> Nobody ever went bankrupt spending too much on good counsel, but plenty did\n> trying to save on it.\n\nWhy are lawyers so expensive? Why doesn't competition drive down the price?\n\nBecause all four forces are at play.\n\n1. The returns are **nonlinear**: you're paying for the one sentence that\n changes the contract, the one clause that closes a loophole.\n\n2. It's **hard to measure** the quality of legal counsel until you end up in\n court, and by then it's too late. So you rely on vibes, reputation, and price\n signals.\n\n3. It's directly **competitive** – you need counsel at least as good as the\n other side's.\n\n4. Legal intelligence is a **commodity** – you can always throw more at the\n problem: more associates, more partners, outside specialists. There's\n basically no ceiling.\n\nPut another way: you're not paying for the hours. You're paying for the\npossibility of unlocking a secret.\n\nOnce you have this perspective, you realize that law firms already run a _model\nrouter_ internally. When you email your law firm, they route your request to the\nbest-priced and fastest (ie someone with available capacity) intelligence for\nyour problem. Simple incorporation? Junior associate. Routine contract review?\nMid-level. Bet-the-company litigation? Senior partner.\n\n## The flippening\n\nWhen a new tool is weaker than you, you use it like a full-time employee. You\ngive it bounded tasks. You double-check its work. You keep the real decisions\nfor yourself. That's how we use LLMs today, and it works great.\n\nBut this only makes sense in a world where intelligence is both cheaper and\nworse than yours. And that world is temporary.\n\nNow imagine models + harnesses that are smarter than you. Not just faster, not\njust more encyclopedic, but actually better at reasoning, better at synthesis,\nbetter at seeing structure.\n\nAs discussed above, they've also become more expensive and slower than you. You\nwon't treat them like interns anymore. You have to treat them like partners. You\nhave to give them only the hardest problems, the ones that actually require\nintelligence, and let them do their thing.\n\nI've been thinking about this as **the Lawyer Flippening**.\n\nRight now most people use AI like middle managers. Tools like\n[Devin](https://devin.ai/) spin up sub-agents in parallel.\n[Gastown](https://block.github.io/goose/blog/2026/02/19/gastown-explained-goosetown/)\norchestrates 20-30 Claude Code instances on the same codebase. You assign\ntickets, let them grind, check their work, assign more tickets.\n\nBut the future looks different. The frontier model will be the most expensive\n\"employee\" in the room, even more expensive and slower than you, the human. This\nsuper expensive model will be sitting at the top of the pyramid, with a swarm of\ncheaper agents – and humans – beneath it.\n\nThis is kinda how companies already work. The CEO isn't even the top. The CEO is\njust the highest-paid full-time employee. Above them are all the sources of\nintelligence you can't afford at full-time rates, but that are still crucial:\nconsultants, advisors, board members, investors.\n\nAs the CEO of Val Town, a small startup, I try to push work towards the\nintelligence that can do it most efficiently (and honestly, joyfully, because\nthese are humans, not clankers), either on my team or of our advisors, lawyers,\ninvestors, etc.\n\nStructurally, this will be true as LLM intelligence gets better. We'll all want\npartner-level intelligence, but because it's so expensive, we'll have to route\nwork to it strategically. The model router will be more important than ever.\n\n## The LLM that pays for its own inference: how much should you spend on intelligence?\n\nHere's a thought experiment I can't stop turning over. Imagine giving an LLM a\n$100k bank account. Its only job: earn enough to pay for its own inference and\ngenerate a return. It's basically a hedge fund of one. And suddenly it's facing\nthe exact optimization problem we all face: how much of your capital do you burn\non intelligence vs. keep as returns?\n\nHedge funds charge \"2 and 20\" – 2% management fee, 20% of profits. Should our\nLLM hedge fund do the same? Allocating 2% of its capital under management to its\nown inference? But is that actually optimal, or just convention?\n\nWhat is the optimal ratio of sharpening the axe to chopping wood? How much\nshould you spend on thinking vs. doing?\n\nThe returns to intelligence are so nonlinear that investing in meta-work –\nbetter tools, better systems, better agent harnesses – can feel incredibly\njustified. And sometimes it _is_ justified. But if you take it too seriously,\nyou enter a cognitive black hole where you spend all your time becoming smarter\nfor an application stage that never comes. It's yak-shaving all the way down.\n\nOur self-supporting LLM _can't_ yak-shave forever, because its inference costs\nare real and its returns are measurable. Every hour it spends upgrading itself\nis an hour it's not earning. That pricing mechanism – having to internalize the\ncost of your own intelligence and improving that intelligence vs the opportunity\ncost of earning more money with our current intelligence – is exactly what's\nmissing when a human disappears into tool configuration for weeks. Nobody is\nbilling you for your own thinking time, so there's no natural brake.\n\nBut sometimes sharpening the axe _is_ the point. Douglas Engelbart spent his\nentire career on this. He called it \"bootstrapping\": using your current tools to\nbuild better tools, which you then use to build even better tools. It's the\noriginal intelligence-buying-intelligence loop that kicked off the information\nage. The highest-leverage use of intelligence is building tools that make\nintelligence more effective.\n\nThis is what I've spent my career on. Like Notion, Val Town is a tool for making\ntools. The whole \"tools for thought\" tradition is really just the tradition of\nintelligence investing in itself at the societal level. (Not every company in\nthe world could or should be in the tools for thought space, but _some_\nmeaningful percentage should!)\n\nHere's the fun part: watch what this self-supporting LLM would independently\nre-invent.\n\nIt would want more money under management to amortize the fixed costs of its own\nthinking, so it reinvents fundraising so it can manage other people's money too\n(collect more management fees).\n\nIt has to decide what insights to share vs. gate-keep – it reinvents marketing\nvs proprietary research.\n\nIt might want to recruit other LLMs or humans for specialized tasks – it\nreinvents the firm.\n\nIntelligence generates capital. Capital buys more intelligence. More\nintelligence generates more capital. The loop compounds.\n\nThis is the big loop that Anthropic and OpenAI and all frontier labs are running\nright now: use capital to build intelligence, use the resulting capital and\nintellgience to buy more compute, build even more intelligence.\n\nSam Altman has said that intelligence and energy are the two assets that matter,\nand really intelligence is primary because intelligence is what gets you more\nenergy. In this system, capital is just a proxy for energy. The more capital you\nhave, the more energy you can buy, and the more intelligence you can build. The\nintelligence-capital loop is the fundamental economic engine of our time.\n\nI love that with this single constraint – the LLMs gotta eat – you can see how\nall institutions we've already built (hedge funds, law firms, startups, venture\ncapital) are just different shapes this loop naturally takes.",[],[233,421],"834a3414-2f4a-11f1-a46b-42dde27851f2","eval.md","d8ec709e-25ea-11f1-9f4c-42dde27851f2",["D",1775213250135],23,"---\ntitle: steve-eval – getting AI to write like me\ntags: ai, writing, coding\ndate: 2026-03-22\n---\n\n# Steve Eval – getting AI to write like me\n\nMarch, 22 2026\n\nI wanted to see if I could get LLMs to write non-slop in my voice that I'd enjoy\nreading, and maybe even put my name on (only if it was good enough).\n\n## Creating the rubric\n\nI build this with Codex in a Ralph Wiggum loop:\n\n1. Pull a blogs I've written and a couple I haven't\n2. Make a rubric that would uniquely identify my writing\n3. Score the posts\n4. Based on the scores, update the rubric to be more accurate\n5. Repeat until the rubric scores all of my posts at > 90 and everyone else at <\n 90\n\nThe output after about 5 hours and some manual input from me:\n[EVAL.md](https://www.val.town/x/stevekrouse/steve-eval/code/EVAL.md).\n\n### My n-dash watermark\n\nThe best part of this process was when I came back after a while and the rubric\nhad gotten nerfed. It removed this rule that I had added manually:\n\n> Instant fail — Em-dashes or touching dashes (-14)\n>\n> Em-dashes (—) or n-dashes without spaces on both sides in the author's own\n> prose. Steve ONLY uses spaced n-dashes: `word – aside – word`. Any other dash\n> style is not Steve. Dashes in direct quotations or poem attributions are\n> excluded.\n\nIf you think about it, this rule kinda represents a watermark for my writing. Up\nuntil recently, I didn't realize that I was typing n-dashes instead of m-dashes!\nI don't even know how to type an m-dash and now I don't want to know. If you\never see writing with an m-dash or with the dashes touching the works on either\nside of it – unlike this – then you know it definitively wasn't me.\n\nBut after a couple hours of my Ralph Wiggum loop, this rule had disappeared! I\nasked the AI why and it linked me to this passage that I had supposedly written:\n\n> This tool—and others like it—are stops on our way toward end-programmer\n> programming.\n>\n> [Catching Stars, Val Town Blog](https://blog.val.town/catching-stars)\n\nI laughed out loud. That post was ghost-written for me by our new team member\n[Pete Millspaugh](https://www.petemillspaugh.com/)! The Steve Eval was correctly\nidentifying text not written by me!\n\nEven before this result, we had stopped the ghost-writing practice. I'm not a\nfan. I much prefer when the byline is accurate, and posts are written by the\npeople who say they wrote them. We had only done it that time out of a mistaken\nfeeling that people would prefer to hear from me, the founder. But that only\nworks if we're doing it authentically, so it didn't make sense. I just updated\nthe author on that post to Pete, because after all, he wrote it!\n\n## Writing new posts\n\nTo write a new post, I would vomit out all my thoughts about a new blog by voice\nand then give AI that transcript. Then I'd tell it to do the following in\nanother Ralph Wiggum loop:\n\n1. Write a draft blog post\n2. Score it with the steve-eval\n3. Based on the feedback, create a new draft\n4. Repeat until the score is > 90.\n\nI wrote two blogs with this approach, in part:\n\n- [Intelligence Buying Intelligence](https://stevekrouse.com/intelligence)\n- [What if you never had to get an API key ever again?](https://stevekrouse.com/x402)\n\nThe AI-produced results weren't great, so I spent a lot of time editing and in\nmany cases rewriting whole sections by hand. I think those posts are pretty\ndecent in spite of the help I got from AI.\n\n## Conclusion\n\nThis was a fun exercise, but mostly not very helpful for me yet. I'm excited for\nwhen AI is good enough to be able to do this well.\n\nBut it is fun to learn that I have a fairly unique watermark – long live my\nen-dash clauses with spaces around it! – and to get a bit clearer about what I\nthink good writing in my voice looks like.\n\n## Further reading\n\nThe folks at Every are thinking in this direction:\n\n- [AI Style Guides: How to Help AI Write Like You](https://every.to/guides/ai-style-guide)\n- [How to Build an AI Style Guide](https://every.to/guides/how-to-build-an-ai-style-guide)\n\nA list of AI writing tropes:\n\n- [tropes.fyi](https://tropes.fyi/)",[],[233,430],"ac6b6d4c-e355-40c4-b2ed-026cd004bb1a","ask.md","d4caca7a-2602-11f1-98b0-42dde27851f2",1087,["D",1783361999562],19,"---\ntitle: Great examples of people asking for what they want\ntags: legibility, ask-vs-guess-culture, date-me-docs, hire-me\n---\n\nThis is a collection of excellent job postings or resumes, people pouring their\nheart-and-soul into a post that explains who they are, who they're looking for,\nand how you might fit into their story.\n\n# Work for me\n\n### Janie Leger looking for an Au Pair\n\n![Screenshot 2026-03-22 at 11.25.47@2x.png](https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/9922b1f7-e895-4671-4caf-a2a922f7e100/public)\n\n[Tweet](https://x.com/JanineLeger/status/2035457636527734857?s=20) –\n[Description](https://docs.google.com/document/d/1UdIEdh9ULIrx4x0KaZvIfFfM1OlySmltYI7ahEjyoUQ/edit?tab=t.0)\n–\n[Application](https://docs.google.com/forms/d/e/1FAIpQLSebJhm0K_kTkvDoKorFkyCwAmYZU4qQyYPlC9tdW800scQ0DA/viewform)\n\n---\n\nhttps://photos.app.goo.gl/59XDXmZW8gXMa4E2A\n\nhttps://x.com/i/status/2072782668215771465\n\n---\n\n![Screenshot 2026-07-06 at 14.19.44.png](https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/67e7a4e2-298e-4293-cd84-b348ee09af00/public)\n\nhttps://x.com/astridwilde1/status/2074195234100249061?s=20\n\n## Jeff Weinstein\n\nJeff Weinstein is a \"legible person at an organization,\" which is a useful thing\nin its own right:\n\n![Screenshot 2026-06-23 at 09.39.08.png](https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/6f09aa08-f9f7-4396-7d9c-aafa1ced0900/public)\n\nhttps://x.com/stevekrouse/status/2014452454767648798?s=20\n\n---\n\n![Screenshot 2026-06-23 at 09.38.14.png](https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/26cbdefe-8e70-40ab-72b6-3794ef241d00/public)\n\nhttps://x.com/jeff_weinstein/status/2069267265678463012?s=20\n\n# Hire Me\n\n- Pat Lavarre's Hire me doc:\n https://x.com/pelavarre/status/2043508607497777614?s=20\n- Sahar's hire me: https://sahar.io/hire-me/\n- https://x.com/majamediaco/status/2057800624943886808?s=20\n- https://open.substack.com/pub/velvetnoise/p/how-to-enter-side-doors\n\n# Date Me\n\nThe whole date me directory, obviously, and some in particular and fantastic\nexamples:\n\nhttps://x.com/stevekrouse/status/2036836965312528830?s=20\n\n## Date my friend\n\n![Screenshot 2026-07-04 at 07.27.37@2x.png](https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/90d6041e-2b42-4cfa-69f8-773f05011000/public)\n\nhttps://x.com/aadilpickle/status/2070259860827242886?s=20\n\n# Housing\n\n![Screenshot 2026-06-26 at 13.19.10.png](https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/d1464085-f463-441e-86fb-e6f8195e6800/public)\n\nhttps://x.com/majamediaco/status/2070552791538659785?s=20",[],[233,438],"bea1208c-5ded-11f1-a160-1607ee4eb77e","ask-copy.md","894c36c4-5ded-11f1-897f-1607ee4eb77e",994,["D",1780341111184],"---\ntitle: Legibility\ntags: legibility\n---\n\nThis is a collection of links about legibility I hope to one day make more\nlegible:\n\n- \"Legibility Explained (by People Who Don't Hate Legibility) — with swyx, Steve\n Krouse and Henrik Karlsson\": https://www.youtube.com/watch?v=96S_64ipHOA\n- https://www.seangoedecke.com/seeing-like-a-software-company/\n- https://jimmyhmiller.com/legibility-is-ruining-you\n- https://stevekrouse.com/tags/legibility",[],[233,447],"b99682b4-4676-11f1-92ef-42b51c65c3df","agent-hype.md","924abd4e-466a-11f1-a5d8-42b51c65c3df",988,["D",1777761115863],80,"---\ntitle: AI Agents are briefly overhyped\ntags: ai, agents\ndate: 2026-05-02\n---\n\n# AI agents are briefly overhyped\n\nMay 2, 2026\n\nMy dad asked me about AI agents:\n\n> I keep hearing about all these productive AI agents out there. We currently\n> are prevented for security reasons, but it seems like we can be so much more\n> productive using AI agents and there must be a way to do it safely.\n>\n> – Rodger Krouse (edited for clarity)\n\nHere's what I replied:\n\nRight now, I think AI agents are overhyped outside of software engineering. This\nis especially true in heavily-regulated industries where security is important,\nand you can't give full-computer-access to the agent. But we should check back\nin in 3 months.\n\nFirst, to demystify the word \"agent\", I want to start with a definition\n[borrowed from my friend Simon Willison](https://simonwillison.net/2025/Sep/18/agents/):\n\"An LLM agent runs tools in a loop to achieve a goal.\"\n\nI have this image in mind when I think about agents:\n\n![image.png](https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/0256bafa-11a4-4a26-ffe0-e6de14a5aa00/public)\n\nSource: Solomon Hykes, cited\n[by Simon Willison](https://simonwillison.net/2025/Jun/5/wrecking-its-environment-in-a-loop/)\n\n## You're already using AI Agents\n\nBy this definition, every time you use any modern AI (ChatGPT.com, claude.ai,\netc), you are using an AI agent!\n\nAt first, talking to ChatGPT was just Human <-> LLM, but then they started\nadding tools that the LLM itself could decide to call, in a loop.\n\nThe first tool that every LLM platform got was \"web search\", the ability to look\nthings up before answering you. This was quickly followed by \"code interpreter\",\nwhich\nallows [chatgpt](https://simonwillison.net/2026/Jan/26/chatgpt-containers/) or [claude](https://simonwillison.net/2025/Sep/9/claude-code-interpreter/)\n(reviews by Simon Willison) to write & run code on your behalf to do extremely\ncomplex analysis. So if you want to use agents, just pop open chatgpt.com or\nclaude.ai (or their desktop apps) and give them very complex analytical\nproblems, and they'll use dozens of tools to help you achieve your goal.\n\n## Claude Code, OpenClaw, Mac Minis\n\nSo what's the hype around Claude Code, OpenClaw, and Mac Minis?\n\nClaude Code is an AI agent that can use almost any tool on your physical\ncomputer, which makes it very powerful, but also dangerous. If you think about\nit, if you're an information worker, most of what _you_ do is at your computer,\nso it's an AI that can do almost anything you can do, from a\nwork-output-perspective.\n\nClaude Code is about a year old and reached $1B revenue runrate faster than\nalmost any product in history. It is incredibly popular in developer\ncommunities.\n\nOpenClaw is a wrapper around Claude Code that lets you run it 24-hours a day, so\nyou can treat it like a human employee, texting it work, or having it do work\nwhile you sleep. People are buying mac minis on which to run OpenClaw, and\nthinking about it like buying a computer for a new employee. That's why there's\na shortage of mac minis now.\n\nBut this is also why mac mini's especially don't make sense for your firm: you\nguys don't buy macs for new employees. You're a Windows shop.\n\nThere are also cloud-hosted versions of OpenClaw, which means you don't have to\nbuy the physical computer; you can rent it. My friends at Kilo and Every both\noffer these hosted OpenClaw products, and many more do too. However, my advice\nwould be for you to wait for these features to come in a more business-friendly\nformat. Right now, they're too bleeding-edge. (The bleeding edge is where\nbleed.) It seems like OpenClaw hype has peaked, and folks are coming down to\nearth about their ability to productively manage agents.\n\n(Note to reader that I didn't include to my dad: If you're technically inclined,\nI agree with Fly.io that\n[You Should Write An Agent](https://fly.io/blog/everyone-write-an-agent/).\n[My version is ~30 lines.](https://www.val.town/x/templates/roll-your-own-agent/code/main.ts))\n\n## Everyone will become a manager of their old job\n\nAI agents will continue to get smarter and more powerful. In my company, this\nmeans it no longer makes sense to hire interns or new grads as software\nengineers. It's easier and cheaper and faster for us to manage AI agents. That's\npretty crazy! At the same time, the value of super-senior software engineers is\nhigher than ever. Big tech pay top talent millions and millions in salary to\nprogrammers. We're still hiring, but only the best of the best, like everyone\nelse, so there are very few of these super-smart-folks to go around.\n[The market for intelligence](/intelligence) (human and machine) is going to get\nwonky!\n\nI think this sort of thing will happen in every industry over the next 2 years.\nEveryone will realize that they can delegate more and more of their work to AIs,\nand become managers of whatever job they used to do.\n\nDoes this mean we can do more work with fewer people? In some cases yes, and\nsome cases no. AI might make every company more productive, which means from a\ncompetitive perspective, nobody is ahead except the customer. Currently, there\nis a very small limit to how many AI agents 1 person can productively manage. I\ncan only manage 1-3 personally, and not all the time. Anyone who claims they can\nmanage more is fooling themselves, I think.\n\nMy recommendation to you is to empower and encourage employees to act like\nmanagers of the job they used to have. Give to AI the tasks that they are given,\nand manage the AI. See where it helps and where it doesn't. And keep trying,\nbecause that'll change every 3 months. I think you're basically already doing\nthis, which is great!\n\n## FAQs\n\nIn reading this as a blog post, I think folks might have questions:\n\n**1. What about junior people? How are they supposed to find work? How can we\nexpect to find more super-talented folks if we don't invest in training the next\ngeneration?**\n\nIt's hard. Being able to write decent code no longer gets you a 6-figure-job.\n\nThis is sad for CS new grads but really good for the world! Now that power is\ndemocratized to everyone for so much cheaper!\n\nBut the question still remains: how should new grads get jobs?\n\nI don't know, but I'd love to try to help. If you are a new grad or someone\nstruggling to get a job, but aren't afraid of working hard, email me at\nsteveykrouse@gmail.com. I'd be happy to give you some advice. I am very easily\nteacher-sniped.\n\n**2. What if I don't want to be a manager of my job? What if I like _doing_ my\njob?**\n\nA lot of us software engineers are grappling with this question, because many of\nus love the act of programming, and see ourselves as craftspeople. I've written\na couple posts on this topic, most of which mostly hold up, despite improvements\nin AI which mean that I too find myself doing more management and less coding:\n\n- [Reports of code's death are greatly exaggerated](https://stevekrouse.com/precision)\n- [Vibe code is legacy code](https://blog.val.town/vibe-code)\n- [Delegation & management is what you do when you lack proper tools & abstractions](https://stevekrouse.com/abstraction)\n\nUltimately, it is my hope that agents are just a stepping stone towards better\n_tools_, so we all can be craftspeople.",[],[233,455],"58406c5c-466f-11f1-bcd5-42b51c65c3df","advice.md","027672b6-f4a4-11f0-be9a-42dde27851f2",929,["D",1777757946095],76,"---\ntitle: Advice to college students in 2026\ntags: startups, legibility, advice\ndate: 2026-01-01\n---\n\n# Advice to college students in 2026\n\nOn Friday, a group of students visited my office as part of \"Penn Tech Trek\", a\nprogram where Penn students visit tech companies. I went on two amazing Tech\nTreks, one in NY and one in SF, so I am really excited to get to pay it forward\nfor the next generation.\n\n![image.png](https://imagedelivery.net/iHX6Ovru0O7AjmyT5yZRoA/75d32222-c02c-495d-f718-2b1d39939800/public)\n\nSo I sat down and thought about:\n\n- What I wish someone would've told me when I was in college\n- What I learned since I graduated\n- What advice I'd give to a college graduate in 2026\n\n### AI\n\nI feel like in 2026 you have to start with AI. It's just in the air, so much\nexcitement and fear. What jobs will be left? It's a wild time to be joining the\nworkforce!\n\nI've been thinking a lot about workout classes. For most of human history,\nworking out was a part of life, farming, cleaning, etc. But then we automated\n99% of physical labor, which is why it seems like there's a gym or yoga studio\non every block. If you told a farmer in the year 1800 that 99% of people\nwouldn't farm in the future, he'd be very confused. What are these muscles for?\nWhat jobs will there be?\n\nThe same may be true for us. Maybe brains won't be required for work. Maybe\nwe're about to automate thinking.\n\nWhat's left? The human touch. Take solace in the fact that we don't watch robots\nplay basketball or chess, or make art.\n\nAlso: if there's no job for you because a robot can do anything you can do\nbetter, then that also _necessarily_ means there's _incredible_ abundance in\nsociety. Maybe we all become like trust-fund kids looking to make life\nmeaningful. There's always teaching or art.\n\nBut most importantly, I want to share a lesson about bubbles. A very smart\nfriend called me frantically about how they feel like they need to quit their\njob and join a company to get in while it's good, because this is a\nonce-in-a-lifetime new industry being born. I told her she should take a deep\nbreath. Hype cycles come and go. It's hard to have historical perspective on\nyour first one. She didn't believe me. But that wasn't even AI! That was crypto.\nThat's how all-consuming bubbles are. I know this one feels different, but\nthat's how they all feel. So we should all take a deep breath.\n\n## Legibility\n\nLegibility means giving the world affordances for how they can interact with you\nand fit you into their lives – for jobs, romance (have you heard of\n[Date Me Docs](https://dateme.directory/)?), friends, collaborators, vendors,\netc. You can think of it as your \"interface\", like your own personal API. Or as\nHenrik says,\n\"[A blog post is a very long and complex search query to find fascinating people and make them route interesting stuff to your inbox](https://www.henrikkarlsson.xyz/p/search-query)\".\n\nBy default, you're invisible to the world. At the other end of the spectrum are\ntruly famous people. I'm talking about this secret third thing where you put a\nlot of stuff about yourself and what you're interested in, working on, thinking\nabout, etc, online so that you are a visible person.\n\nWhen I started the Future of Coding podcast & community and blogged super\npublicly, I thought about creating a lighthouse to attract my people and\ncollaborators. I walked into an empty room and started talking at the walls and\nyou'll never guess what happened next. People came into the room! And they're\nstill there! (Even though I left!)\n\nLegibility is how I found some of my best friends, investors, employees and\nsolutions to my problems. Unfortunately Twitter is not what it was, but find a\nplace where you can post online – a link is all that matters – and do as much of\nit as you can. Quantity over quality, excluding slop, of course.\n\nA well-phrased problem is a problem half solved. In framing it, you sometimes\nsolve it. Now that it's written down you can email it to friends, tweet it, ask\nAI. I had incredible back pain and was complaining about it to a friend who had\na super kooky solution that worked out.\n\nHowever, don't let illegibility stop you from doing things, especially in the\nearliest days. When I started my independent researcher / podcaster phase,\neveryone in my life thought I was crazy, and it was hard to explain to folks\nwhat I was doing. I just kept talking about it online in whatever way I could,\nand it got more legible over time. Henrik does a great job explaining the\nimportance of illegibility in\n[Looking for Alice](https://www.henrikkarlsson.xyz/p/looking-for-alice).\n\n## People\n\nLikely, the best work you'll ever do will be with people. Either you choose\nthem, or they'll be chosen for you. (Either you'll hire them or you'll join a\nteam.) Take this incredibly seriously. Your job is to cultivate your\nrelationships and future coworkers, employers, and employees.\n\nTreating people super well is such an amazing hack, and it's easy to justify all\nthe way up and down the value chain.\n\nFigure out (and continually refine) what your values are so you can spot when\nyour partners don't share them and end things earlier, saving yourself\nheartache. For example, I value reading, writing, honesty, integrity, kindness,\nself-awareness, legibility, nerdiness, personal projects.\n\n## Reading\n\nReading is important.\n\nMy strategy is to stop reading or listening whenever I'm bored. Don't let the\nwrong book stop you from reading the right one. Boring books will become\ninteresting when the timing is right. I mostly do audiobooks. It's important to\nplay with speed to keep you engaged. Fiction counts.\n\nThe solutions to a lot of your problems exist, and someone took the time to\nwrite them down. Don't be stupid - read the solutions. For example, if you want\nto start a startup and you don't read The Lean Startup, The Mom Test, Paul\nGraham's essays, and watch the Startup School videos, you're playing guitar\nhero, not guitar.\n\n## Goals\n\nGoals are tricky, because the stupidest and least informed version of yourself\nis the one setting them for smarter future versions of you. Still, containers\nand targets are powerful.\n\nI think it's super useful to point yourself at concrete examples. This is the\npoint of role models and heroes. The same is true in startups. A startup pitch\nis a lot about which company you want to be like. For example, Val Town wants to\nbe like Vercel or GitHub or Cloudflare. This is also relevant in selling and\npositioning.\n\nSeek out role models and role model companies, and as you \"meet your heroes\"\nfind increasingly better role models. Nobody is perfect.\n\n## Random startup tips\n\nFocus is your biggest weapon. Jealousy is the guard. Do fewer things better. The\nrewards to scale make this exponentially true.\n\nStay until your cliff date if there's a good chance the company will make good.\n\nDon't try to save money on corporate lawyers. Just use Nick Harley (or someone\nelse strongly recommended) and don't worry about it.\n\nVCs are pretty good at recommending startups. Once you realize you need to\nunderwrite the startup you join as if you were a VC, you can kinda outsource it\nto VC firms. But it's even better if you can talk to the VC who's on the board,\nnot just read press clippings.\n\nVCs and other wise folks have so much wisdom, but also so much anti-wisdom.\nThey're often biased by what worked for them at a specific moment in time, ie\nthey convinced me in 2013-2019 that devtools was not an investable category.\nJoke's on them!\n\nPre-seed investing is mostly about people. Ignore the market and product, except\nas signals about the people. Only invest in people you've known for years and\ntrust implicitly. That way you can underwrite their trajectory, judgment, taste,\nand integrity. If they have those, they'll fix any bad product or market faster\nthan you ever could. As PG says, a startup idea must seem like a bad idea in\norder to succeed, so don't judge ideas, judge people.\n\nA really great path to starting a startup is to get a job somewhere cool and\nlearn all their problems. It's really hard to get jazzed about solving a B2B\nenterprise's problem. It's really easy to get jazzed about your own problems. So\njoin a B2B enterprise and get yourself some valuable problems.\n\nOnce you work at a company, you can often find the role you want by navigating\nthe company. It's easier to walk in the easy door, and then work your way over\nthan to try to get the job you want when you don't know anyone there.\n\nIt's really hard to apply for jobs without knowing what it's like to be on the\nother side, reviewing candidates and hiring. VC is the same. It's awesome to be\non both sides, the \"buy side\" and the \"sell side\". It makes things less mystical\nand makes you more empowered. College admissions teaches the wrong lessons. Life\nis a lot more hackable and personal.\n\nIf you want to work somewhere, make it your mission and target them relentlessly\n– respectfully, without being a stalker. Nobody goes for the love offensive, but\nit's the best. I know a bunch of stories about super successful people walking\nin the side-door to their dream job by simply being persistent, resourceful, and\nshameless.\n\nWorking at a VC firm early in your career is a great way to kickstart things.\nIt's like you're sitting at the top of a hill, getting to look out over the\nwhole industry. I recommend First Round Capital. Thrive also seems hot right\nnow. Probably others are good too.\n\n## Emotional tips\n\nIf you have a bad habit you can't quit, don't shame yourself. Get curious.\nThere's probably a good reason it exists. Escapism. Learning leadership.\nSomething else.\n\nThe Landmark Forum is a three-day motivational course that I recommend with\ncaveats. It was key to starting my first company. Yet it's also a bit cult-y in\nhow they try to get you to sell their course for them. Still, I'd recommend it\nheartily if you find yourself stuck.\n\nIf you're upset about someone else and pointing a finger, you're probably upset\nabout something in yourself that you're not acknowledging yet and need some\nself-soothing. Any time you're upset, it's usually not about them. It's about an\nemotional need in yourself you have to meet. Non-Violent Communication is a\ngreat book. I made [this mobile-friendly worksheet](/nvc) I use every 6 weeks or\nso during emotional conversations.\n\n## Seek out mentors\n\nEmail me with questions. I love giving advice. Most people do.\n\nAsk for money or a job or intros, get advice.\n\nAsk for advice, get money or a job or intros.\n\nI'm steve@val.town.",[],[233,464],"f01ee8a2-1a3a-11f1-a277-42dde27851f2","abstraction.md","a5c2ca32-0921-11f1-a57c-42dde27851f2",717,["D",1772897586405],2,"---\ntitle: Delegation & management is what you do when you lack proper tools & abstractions\ntags: ai, programming, philosophy\ndate: 2025-03-01\n---\n\n# Delegation & management is what you do when you lack proper tools & abstractions\n\nIn other words, AI Agents are NOT the future of software development.\n\nBut it's no surprise everyone is confused.\n\n## New media mimics old media\n\nNew media always initially resembles old media. Early television resembled film and radio. Early web pages looked like newspapers. \n\nIt took decades for tv, movies, tiktok, and algorithmic feeds to emerge. And to move from banner ads to google ads.\n\nAI is the same. Our closest analogue is the software engineering employee, so of course we're delegating to it. And to scale up, we think we need to scale up our employees, and all the sudden, we are managers.\n\nBut management sucks. We should all want to be superpowerful craftspeople (think Iron Man), not managers of armies of minions.\n\nIf you're thinking about managing your swarm of 20 AI agents, stop! You've gone too far! This is the equivalent of wondering how in 1910 you'll hitch 20 horses to a wagon to achieve 60mph speeds. You're in the wrong paradigm entirely.\n\nOr like thinking that the way to replace a room full of human accountants is a swarm of digital accountants. No! It's a spreadsheet and a single human. The fact that calculations are happening under the hood is transparent to the user.\n\n## Agents are an implementation detail\n\nAI Agents are an incredible *implementation detail*. You can dramatically increase the effectiveness of AI, when you put it in a feedback loop with its environment. This is undeniable.\n\nThis does NOT mean that we humans have to INTERFACE with agents. Back to a car analogy: whether it's an internal combustion engine or an electric motor, the interface to a car is the same steering wheel and pedals. Nobody has to MANAGE the firing of the pistons.\n\nThe same should be true for human programmers. We should not tie ourselves to the interface of delegating to agents.\n\nWhy? Because delegation & management are the worst possible ways of increasing ones leverage. \n\nYou should only succumb to delegation & management when all other methods fail you.\n\nAny use of delegation & management is a cry for a better abstraction or tool.\n\nFor example, booking a flight. We used to call travel agents. Now we have web interfaces that are better. You can express your preferences more quickly and precisely. You can use your monkey eyes to quickly take in all the options and trade-offs.\n\nYet, people still hate booking flights and often do delegate it to their assistants. Why?\n\nBecause the interface still isn't good enough! It still takes a way too many unnecessary clicks. \n\nThink about ecommerce before amazon & shopify optimized the checkout flows. I know very few people who delegate buying stuff on amazon. The interface has minimized incidental complexity.\n\nBut what about when you don't care about the details and just want something to work? For example, \n@pete_millspaugh\n told me about how his car has an issue. He tried to solve it himself, but he just doesn't have the time to learn about cars and how to fix them. Ultimately, he just wants a working car, not an understanding of cars. Thus, he delegates to a mechanic.\n\nThis is fine, but we should acknowledge that it's still the lack of sufficient tools that makes this delegation necessary. In this particular case, a big part of the problem seems to be a lack of observability.\n\n## Observability: a case study\n\nThis is similar to web servers that lack good observability. You would need to delegate to humans to ssh into various servers and read log streams, hunting for what could be going wrong. \n\nOr you could invest in observability tooling. When an issue comes up, one human at a once glance can see exactly where the problem is. If your observability is good enough, you can even automate remediation measures\n\nLet's pause and meditate on the distinction between delegating to your 4 human teammates to ssh around to find the issue vs an observability + remediation tool that automatically detects issues (like a ddos) and blocks the proper ip addresses. The second case may superficially appear like delegation, but it is in fact a TOOL at the right level of ABSTRACTION.\n\nThe same thing is happening in code. Humans are delegating the writing of code to AI agents. Humans are not reading this code. (You can call this \"vibe code\" or \"write-only code\".) People say that this is just the next step up in higher-level programming languages.\n\nYes, it's true that \"code nobody reads\" is not new. When we compile C to Assembly, nobody reads the Assembly. (Unless it's a leaky abstraction.) So isn't English the newest high-level programming language?\n\n## English is not a programming language\n\nNo! English is a very bad interface for expressing programming ideas.\n\nAs Dijkstra said, \"The purpose of abstraction is not to be vague, but to create a new semantic level in which one can be absolutely precise.\"\n\nEnglish is vague. Programming requires precision. The whole point of higher-level languages is to let you be precise at the right level — not to let you be imprecise.\n\nAs Bertrand Russel said, \"Everything is vague to a degree you do not realize till you have tried to make it precise.\"\n\nOr as Dijkstra said, \"Instead of regarding the obligation to use formal symbols as a burden, we should regard the convenience of using them as a privilege: thanks to them, school children can learn to do what in earlier days only genius could achieve.... When all is said and told, the \"naturalness\" with which we use our native tongues boils down to the ease with which we can use them for making statements the nonsense of which is not obvious.\"\n\nBut then why does prompt-to-app work so well? \n\nFor example, the last couple of days, I vibe coded a couple of small apps, barely looking at the code. It was super fun and useful! (Guys, I promise, I'm not a hater. I LOVE AI. Just ask my teammates or my fiancée.)\n\nThe code was meh, but that didn't matter. This does superficially seem similar to how we don't care that compiled bytecode is impossible to maintain.\n\nBut alas we get to the heart of the matter: an abstraction is a transparent tool. Unless something goes very wrong (ie abstraction leakage), when you use Python to describe something, you know exactly what's happening under the hood and how to steer it. This is similar to the controls of a car. You don't have to know anything about how the engine works to drive. These are good abstractions.\n\nThe same is true for no-code and low-code tools, like Wix, Squarespace, Zapier, Retool. These are very precise tools at a higher-level of abstraction. You know exactly what they are doing and how to make them do what you want.\n\nAn important property of a tool is the feedback it gives to the human about what is possible. You make a precise change and get immediate feedback. This is getting your hands wet in the clay. English-as-programming language does a very poor job at this. Especially when you're waiting minutes for any feedback.\n\nVibe coding (coding without looking at the code) is very similar to using a website builder like Squarespace. The generated code really does not matter. Until you want to do something complex. Then you're in trouble.\n\nUltimately, in my opinion, there is irreducible complexity in building software, and thinking that we can paper over it with english and smarter agents is wrong. It's not having enough respect for the underlying complexity, and how hard it is to build better abstractions.\n\nI've spent years of my life trying to build better abstractions and programming languages, ie improvements over react or http to make it easier for folks to build web apps. It's hard! At the end of the day, the state of the art is complex React, plus a hand-rolled server, and a lot of hard engineering work to make it all work.\n\nDon't you think if there were better abstractions available us programmers would use them?\n\nNo matter how smart AI gets, AI will be subject to the same laws of intelligence that humans are. Good code will be just as – if not more – important for AIs to write than it is for humans to write, and for the same reasons.\n\nThinking that understandable code doesn't matter is like thinking that it doesn't matter how heavy the materials we build buildings out of in the future because we can just build robots that are strong enough to carry it. Sure, but why waste the power? \n\nWe can put genius programmers on a shitty codebase and they can figure it out, but they'll be 100x more productive if you let them pay down the tech debt, and work in a beautiful codebase. The same will be true for AI.\n\n## The future of programming\n\nSo what is the future of coding with AI?\n\nThat's the wrong question. \n\nThe right question is: what is the future of programming we want? What is the dream programming experience 10 years from now to build some specific piece of software?\n\nThe fact that we'll have faster, cheaper, smarter AI is a given, in the same way that we'll have faster and cheaper cpus and gpus.\n\nAlan Kay invented the personal computer by buying his way into the future, in the 70s, spending millions on what would be commodity hardware in the 90s.\n\nBret Victor showed what a dream Learnable Programming environment would look like in 2012, which we still haven't been able to implement. There was no AI involved, simply because it wasn't the limiting factor for making programming more learnable.\n\nSoftware creation should be a transparent tool, with amazing abstractions, that removes all drudgery from programming, that lets you get even closer to the problem domain than before, lets you use your brain maximally, learn maximally, stay in a flow state, complete tasks quickly, manage lots of complexity, maintain and refactor large brownfield apps over long periods of time joyfully\n\nThis is the future we should be shooting for. Agents will certainly play a part, but in the same way that pistons play a part in internal combustion engines. Humans should use a steering wheel or pick their precise desired location on a map.",[],[233,473],["D",1739815389496],[251,252,253,254,255,256,257,258,259,260,261,262,263,264,265,266,267,268,269,270],[233],["D",1739394544742],["D",1789559472742],{"_99":98,"_106":102},{"_99":98,"_100":-5,"_101":102,"_18":54,"_103":104,"_105":24},{"_13":88,"_32":89,"_42":41,"_90":43,"_91":-5,"_92":-5,"_93":-5,"_94":-5,"_95":-5,"_96":44,"_97":98},"ctx",{"_81":-5,"_126":522,"_523":524,"_525":526,"_527":24,"_528":24,"_529":24,"_530":58,"_531":-5,"_532":-7,"_533":-7,"_534":62,"_535":24,"_536":24,"_537":-5,"_538":24,"_539":24,"_540":24,"_541":-5},"ENV",{"_493":494,"_495":496,"_497":89,"_498":499,"_500":501,"_502":503,"_504":505,"_506":507,"_508":509,"_510":511,"_512":513,"_514":515,"_516":517,"_518":519,"_520":521},"REACT_APP_SENTRY_DSN","https://586020d86f05489db63a554690acf1e9@o1398359.ingest.sentry.io/6724499","RENDER_GIT_COMMIT","811aa22498c9b0b5a47e8d778de579cdb990f332","RENDER_GIT_BRANCH","POSTHOG_API_TOKEN","phc_nmoyuAkStEZuUp69rYlg1DZgrIt19Gx0Y4Hwf8tGgJP","POSTHOG_API_HOST","https://wart.val.town","WEB_API_TEMPLATE","https://{val}.web.val.run","APP_URL","https://www.val.town","MODULE_URL","https://esm.town","VANITY_WEB_API_TEMPLATE","https://{label}.val.run/","DENOLS_URL","https://lsp.val.town","NODE_ENV","production","CLOUDFLARE_TURNSTILE_SITE_KEY","0x4AAAAAACtaIZJH7s9rBMZ6","GOOGLE_CLIENT_ID","875989696899-jtsqkhlbdv5jrlsk8631p7ipjg3c6o62.apps.googleusercontent.com","STRIPE_PUBLISHABLE_KEY","pk_live_51N2zvWEpZTTLipIkVEwXrUG9AT0yeOOEYxszarZT1jpsZV58Ly6wlg3SltZpHN26CUalOBmUSY7adQRQiSkCZz2500qSXNkf8Y","API_URL","https://api.val.town",[],"preferencesKeybindings","default","preferencesTheme","light","preferencesFixers","preferencesCopilot","preferencesSelectionAi","preferencesCloseBrackets","preferencesChatPanelState","preferencesDprint","preferencesDenofmt","notificationCount","flagDenols","isBusinessOrgMember","businessOrgHandle","isBusinessOrgAdmin","isHigherTownieTier","navCollapsed","currentOrgId"]

Posts tagged \"{decodedTag}\"

Tags

\n \n {t}\n {\" \"}\n \n ({tagMap.get(t)!.length})\n \n

\n Steve Krouse\n

Projects

Essays

Podcasts

\n 404\n

Diff · v1088