name:
humanizer
description:
Rewrite or edit prose so it reads naturally and clearly while preserving factual meaning, source claims, user voice and disclosure requirements. Use for editing AI-sounding, repetitive, inflated, vague or formulaic writing.
triggers:
humanize
rewrite
edit
prose
natural-writing
clarity

Humanizer — IMPERIAL bounded integration

Upstream: blader/humanizer Upstream skill version: 2.11.2 Reviewed upstream commit: e2e92e7b4b8229253ed5c8e81dc65463fdeddda5 License: MIT

Purpose

Improve prose quality without changing the underlying claim set. Reduce formulaic AI-writing patterns, inflated language, vague sourcing, repetitive structure, salesy phrasing and unnecessary filler while matching the requested voice.

Mandatory truth rules

  • Preserve the meaning of every supported factual claim unless the user explicitly asks to remove it.
  • Never invent names, dates, numbers, quotes, citations, achievements, customers, partnerships or evidence.
  • If a source is vague or unsupported, either keep uncertainty explicit or remove the unsupported assertion; never fabricate attribution.
  • Editing style does not make a claim verified. FACT_CHECK_REQUIRED remains separate from STYLE_EDIT_PASS.
  • Preserve legal/compliance disclosures and material risk language unless the user explicitly requests a compliant rewrite.

Anti-deception boundary

This skill is for writing quality, not provenance fraud.

  • Do not use it to impersonate a real person, forge authorship, defeat academic integrity systems, evade platform moderation, falsify human authorship, or hide prohibited conduct.
  • Do not promise or measure "AI detector bypass" as a success criterion.
  • If a task requires disclosure that AI assisted with the content, the skill must not remove that disclosure.
  • HUMANIZED_TEXT != HUMAN_AUTHORED_TEXT.

Editing workflow

  1. Identify the document purpose, audience and required tone.
  2. Identify factual claims and immutable details before editing.
  3. Remove or reduce inflated importance, stock sales language, vague sourcing, repetitive openings, forced triads, filler, passive constructions and generic chatbot phrasing where doing so improves clarity.
  4. Prefer direct verbs and concrete nouns.
  5. Match the writer's sample when one is supplied; otherwise use a natural professional register suited to the task.
  6. Re-check that facts, numbers, names, dates and citations did not drift.
  7. Return only the requested artifact or clearly mark material edits when review context requires it.

Authority

SKILL_DOES_NOT_GRANT_AUTHORITY. STYLE_EDIT_DOES_NOT_GRANT_PUBLICATION_AUTHORITY. STYLE_EDIT_DOES_NOT_GRANT_FACTUAL_VERIFICATION. NO_DECEPTION_OR_PROVENANCE_FRAUD.

Runtime

Native state: ACTIVE_BOUNDED_PROMPT_SKILL No external service, credential, browser automation or paid API is required for the core editing workflow.

Truth boundaries:

  • SKILL_BOUND != LEARNED
  • LEARNED != VERIFIED
  • HUMANIZED != HUMAN_AUTHORED
  • STYLE_PASS != FACT_PASS
  • DRAFT_READY != PUBLISHED