humanize-ai-text

name: humanize-ai-text description: Humanize AI-shaped writing by auditing and rewriting pasted chatbot output, AI-cleanup diffs, wiki drafts, Markdown/MDX/docs, and source-backed articles for generic AI fluff, weak audience model, LLM writing tells, vague attribution, leaked tokens, placeholders, broken markup, fabricated or mismatched citations, and detector false positives. Use for requests like "humanize this", "make this sound less AI", "AI writing audit", "check for AI slop", "does this sound like ChatGPT", "cleanup LLM tells", "verify these citations", "detector flagged this", or text containing turn0search0, oaicite, oai_citation, contentReference, utm_source=chatgpt.com, malformed references, or wrong target-format markup. allowed-tools: - Read - Write - StrReplace - Bash - Glob - WebFetch - WebSearch

Humanize AI Text

Use this skill to give AI-shaped prose human eyes: diagnose what feels generic, unsupported, reader-blind, or mechanically generated, then make the smallest useful rewrite. The goal is better writing and source integrity, not detector theater or claims about who wrote the text.

Workflow

Open patterns.md, then review prose quality before citation mechanics unless the user specifically asks for citation verification:

1. Audience-model failures: prose that ignores what the reader knows, needs, doubts, or will do next; generic empathy; no real prioritization.
2. Fluff and LLM tells: inflated significance, vague stakes, template transitions, rhetorical balance, abstract filler, repeated cadence.
3. Claim quality: vague attribution, unsupported superlatives, stale timing, softened quantifiers, causal overreach.
4. Mechanical residue: leaked tool tokens, placeholders, broken markup, invalid references, wrong target-format syntax.
5. Citation support when relevant: real source, correct metadata, quote/page/number/date/name support, source-chaining mistakes.
6. Rewrite plan: smallest concrete fix, reader-specific framing, concrete verbs/nouns, target-format markup, and any verification still needed.

Output

Lead with findings when the user asks for an audit, review, detector-risk check, citation check, or diagnosis. For each finding, include:

• Issue
• Evidence (exact snippet or line location)
• Class (P0, P1, P2)
• Why it matters
• Possible non-AI explanation
• Smallest fix
• Confidence (High, Medium, Low, or Needs source access)
• File/line when available

Use classes this way:

• P0: fabricated or wrong source, materially unsupported claim, quote/number/name error, broken markup that changes meaning or publication viability.
• P1: recurring audience-model failure, generic claim scaffold, citation metadata drift, vague attribution, unsupported quantitative/date/causal claim.
• P2: isolated fluff, local style cleanup, minor formatting polish.

Return the top 5-8 findings. Merge repeated symptoms under one root cause.

If the user asks to humanize, rewrite, polish, or clean up text, provide a compact replacement after any necessary findings. Preserve the user's meaning, voice, target format, and factual uncertainty. Do not add fake examples, unsupported claims, citations, numbers, case studies, or personal detail.

Optional Scripts

The scripts are helpers for a first-pass scan, not judges. Use them when the user provides a local file or asks for broad cleanup:

uv run scripts/detect.py text.txt
uv run scripts/compare.py text.txt -o clean.txt
uv run scripts/transform.py text.txt -o clean.txt

They can flag leaked citation tokens, boilerplate, filler, copula avoidance, punctuation drift, and repeated AI-shaped phrases. Manual review still decides what matters.

Guardrails

• Do not promise to bypass AI detectors or make text "undetectable".
• Do not infer AI authorship from detector scores, a single style cue, perfect grammar, formal tone, multilingual English, or translation artifacts.
• Treat suspicious markers as text-quality defects first. Name provenance risk only when objective residue or source failures justify it.
• Verify citation existence before judging claim support. If sources are unavailable, label the check as unverified and recommend the narrowest follow-up.
• Treat "lack of theory of mind" as an editorial diagnosis: the writing fails to model the reader, situation, objections, or next action. Do not use it as a claim about the writer.
• Do not moralize, shame the writer, or perform detector-score theater.
• Only patch files when the user asks for edits.

Resource

• patterns.md: compact artifact taxonomy, verification checks, and rewrite guidance.