| 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
- Read the user text and identify its target context: personal post, essay, wiki/article draft, technical doc, comment, or marketing copy.
- 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.
- 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.
- Load
references/signs-of-ai-writing.md when you need the detailed checklist, especially for transition bloat, repetitive contrast patterns, or image-caption redundancy.
- 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.
- Report only actionable findings. Quote short snippets from the user text, explain why each feels synthetic or weak, and give a concrete replacement.
- 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.
- 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.
