audit-ai-writing

name	audit-ai-writing
description	Audit and revise drafts for likely AI-writing tells using the Wikipedia "Signs of AI writing" field guide. Use when the user asks to check, review, humanize, de-AI, polish, rewrite, or make writing sound less AI-generated; useful for essays, posts, documentation, comments, and article drafts.

Audit AI Writing

Use this skill to find and fix AI-writing signals without pretending to be an AI detector. Treat the signs as review heuristics, not proof that a text was generated by AI.

Workflow

1. Read the user text and identify its target context: personal post, essay, wiki/article draft, technical doc, comment, or marketing copy.
2. For public drafts, first check the audience boundary: every sentence in the published text should address readers, not the author, editor, Codex, Claude, an agent, or a reviewer.
3. For illustrated public drafts, do an image-text economy pass: if an image already shows the action, evidence, UI, or joke, the surrounding prose should add context or judgment, not repeat what readers can see.
4. Load references/signs-of-ai-writing.md when you need the detailed checklist, especially for transition bloat, repetitive contrast patterns, or image-caption redundancy.
5. Separate issues into four buckets:
   • Content problems: vague significance claims, unsupported notability, generic analysis, promotional tone, missing specifics.
   • Language problems: canned AI vocabulary, negative parallelisms, rule-of-three rhythm, over-smooth transitions, weak attribution.
   • Format/source problems: Markdown artifacts, excessive bold, em dashes, broken citations, fake-looking references, placeholder links.
   • Audience-boundary problems: editor notes, model self-talk, instructions to the author, or questions to the author that leaked into reader-facing prose.
6. Report only actionable findings. Quote short snippets from the user text, explain why each feels synthetic or weak, and give a concrete replacement.
7. If the user asks for a rewrite, preserve their intended meaning and voice. Prefer deleting filler, collapsing redundant transitions, and adding specificity over merely swapping synonyms.
8. Do not accuse the author of using AI. Say "this reads like an AI-writing tell" or "this pattern can feel generated".

Output Styles

For a review, use:

## 主要问题
- `snippet` -> 问题：... / 建议：...

## 可直接替换
...

For a rewrite, provide the revised text first, then a short note listing the largest changes.

For short text, keep the answer concise. For long text, prioritize the top 5-8 highest-impact edits instead of exhaustively marking every phrase.

Guardrails

• Do not rely on AI detection scores or claim certainty about authorship.
• Do not remove all personality; some informal repetition, asymmetry, and directness can make writing more human.
• Do not over-polish into corporate blandness.
• Do not preserve explanatory bridges just because they are logically correct; public drafts often improve when obvious transitions are deleted.
• Keep source/citation checks strict: fabricated or mismatched references are higher risk than surface style.
• Keep publication boundaries strict: rewrite or remove instructions that were meant for the author, editor, or AI assistant before producing reader-facing text.