--- name: ai-writing-audit description: >- Audit a document for the signs of AI authorship and rewrite to remove them. Judges by clusters and structure, not a single banned word: leaked citation markup and low sentence-length variance are near- decisive; negative parallelism, formatting overkill, and compulsive summaries fire when they co-occur; vocabulary and em-dashes only contribute. Use when a draft, report, brief, handover doc, marketing copy, README, or any prose you are about to share needs to read as a person wrote it rather than an LLM. Triggers include "audit for AI tells", "remove the signs of AI writing", "make this sound less like AI", "de-slop this", "proofread for AI patterns", and references to the Wikipedia "Signs of AI writing" essay. --- # AI writing audit Detect the tells that mark prose as LLM-written, then rewrite so it reads as human. The pattern catalogue, with examples and sources, is in [`REFERENCE.md`](REFERENCE.md); the triage scanner is [`tools/audit.py`](tools/audit.py). This page is the method. Last verified against the 2026 landscape (see *Provenance* at the end). ## When to use this skill Apply it whenever you (the assistant) have produced, or are about to produce, prose a real person will read as human-written: handover docs, briefs, decision docs, marketing and landing copy, social posts, emails sent on someone's behalf, READMEs, developer-facing prose. The user can also invoke it directly ("audit this for AI tells", "de-slop this draft"). ## The one principle: judge clusters and structure, not single words The field moved on from word-lists. A lone "crucial" or a single em-dash proves nothing, and the loudest old tells have been trained or instructed away (frontier models dialled back "delve" through 2025, and OpenAI shipped an em-dash-suppression setting in late 2025). What survives that arms race is not any one word. It is **how many tells cluster in one place** and **the shape of the writing underneath**. So weight the evidence in three bands: - **Near-decisive on its own.** Leaked citation or tool markup (`oaicite`, `[cite: 1]`, `turn0search0`, `grok_card`, `ppl-ai-file-upload`). These are machine artifacts, not style; one is enough to conclude the text was pasted from an LLM. Very low sentence-length variance (robotic rhythm) is the strongest *stylistic* signal. - **Fires when it co-occurs.** Negative parallelism ("it's not X, it's Y"), formatting overkill, the rule of three, compulsive summaries, trailing "-ing" analysis, false ranges, praise-first sycophancy openers. One is ordinary human writing. Three or more stacked in a short span is the tell. - **Contributes only.** Vocabulary (era-tagged in REFERENCE.md), hedging clichés, stock transitions, flattery adjectives, em-dash *density*. Never conclude from these alone; count them toward a cluster. Rewrite to reduce **density across bands**, not to score zero on any single pattern. A human wrote "crucial" too. ## The headline checks Full detail and examples live in REFERENCE.md. In brief: 1. **Citation / markup artifacts** (near-decisive). Search for leaked tokens: `oaicite`, `contentReference`, `turn0search0`, `[cite: 1]`, `[span_1]`, `grok_card`, `ppl-ai-file-upload`. Delete them and verify the claim they were attached to. 2. **Sentence-length rhythm** (strongest stylistic). LLM prose clusters near one length with a mid-sentence comma; human prose varies far more. Read aloud. Break the monotony with short sentences and the occasional fragment. 3. **Negative parallelism** (now the most common single tell). "It is not a product, it is a movement." Rewrite to a direct claim. 4. **Formatting overkill.** Bold on most lines; a bulleted list whose bold title the next sentence restates; decorative arrows and emoji headers. Use bold for headings and load-bearing terms only. 5. **Rule of three.** Stacked triplets where the third item is filler. Vary the cadence: two items, or four. 6. **Compulsive summaries.** "In conclusion", "Overall," on short sections that need no summary. Cut them. 7. **Trailing "-ing" analysis.** "Sales rose 12%, reflecting strong demand." Split into two sentences or cut the tail. 8. **Sycophancy openers** (a 2026 addition). "Great question", "You're absolutely right", "Let's break this down." Delete; answer directly. 9. **Vocabulary, hedging, transitions, flattery, em-dashes** (contributory). Suspicious in a cluster, fine alone. See the era-tagged list in REFERENCE.md before flagging a word. ## How to apply 1. **Scan for triage.** Run the scanner against the document or its source text: ``` python3 tools/audit.py path/to/file.md ``` It accepts plain text, Markdown, Python (it reads string literals), and PDF (via `pypdf`). It groups hits by evidence band and estimates a cluster verdict. Treat the output as a starting point, never a ruling: it cannot judge rhythm, whether a triplet is brand voice, or whether a bullet title is load-bearing. 2. **Read it yourself.** The scanner misses the two strongest signals almost entirely: monotone rhythm and formatting overkill in tables. Read the document aloud. If a sentence would not come out of a person's mouth, mark it. 3. **Rewrite, do not surface-edit.** Deleting the offending word leaves the shape intact. Rewrite the clause. Common moves: negative parallelism to a direct claim; a triplet to two items or a plain sentence; a buzzword to a plain verb; a monotone run broken with a short sentence. Reduce the *cluster*, not the single hit. 4. **Keep the author's voice.** Phrases from a brand voice guide, a bio bank, or a real human author are not yours to flatten just because they pattern-match. The job is to remove tells that came from the model, not to sand down a person's writing. When unsure which phrases are intentional, ask. 5. **Re-scan and confirm.** Run the scanner again. Confirm the near-decisive band is empty (no leaked markup), the cluster count dropped, and no band is stacked. Report what changed. ## Output When asked for an audit, produce: 1. A short summary: which tells appeared, in which evidence band, and the cluster verdict. 2. The rewritten document, or a list of suggested rewrites if the user only wants a review. 3. The post-rewrite scan. Do not narrate the rationale sentence by sentence. The user can read the diff. ## Provenance The catalogue tracks the Wikipedia "Signs of AI writing" essay and 2025–2026 primary sources (system cards, stylometry studies, the tropes.fyi catalogue); REFERENCE.md carries the citations. **Last verified 2026-07.** The vocabulary layer is the most perishable part of this skill: individual words get patched out within a release cycle once they become notorious, so re-verify the era-tagged list before trusting it. The structural signals (rhythm, clustering, leaked markup) age far more slowly. Update REFERENCE.md, not this page, when new patterns surface.