AI Tell Detector & Rewriter

What it does

It turns "this sounds like AI wrote it" into a number. The skill scores a draft against 8 concrete, nameable tells — empty intensifiers, too few verifiable numbers, no first-person perspective, over-uniform sentence and paragraph lengths, symmetric triads, transition-word dependency, no cost or uncertainty admitted, and high abstract-noun density.

Each tell contributes weighted points, so you get a 0-100 score before you touch a word, plus a list of which tells fired.

Score	Reading
0-20	Reads human. Leave it alone.
21-45	Traces present. One editing pass.
46-70	Clear AI smell. Fix item by item.
71-100	Rewriting is cheaper than patching.

Then it rewrites by six methods, each chosen to match the tells that actually fired. Diagnose first, then rewrite — rewriting without diagnosing is putting makeup on in the dark.

Why use it

Detection systems on Chinese content platforms (WeChat Official Accounts, Zhihu, CSDN, Xiaohongshu) flag "low originality / suspected AI" and cut distribution. Writers who rewrote by feel get rejected anyway. This skill makes the diagnosis explicit and repeatable, so the rewrite has a target.

The same reasoning applies anywhere an audience has learned to smell machine text: the tell list is about structure and specificity, not vocabulary.

How it works

1. Paste the draft. The skill scores all 8 tells and prints the 0-100 number with the fired tells and their weights.
2. It names which tells fired — not "make it more human".
3. It applies the matching methods from a six-method set (concrete detail substitution, perspective injection, asymmetry, trade-off disclosure, and two more).
4. Ships with a runnable checker (scripts/ai_trace_check.py, pure Python, no dependencies, no network) so the score is reproducible instead of a vibe. Reads a file, stdin, or --text. --json for piping.

Language support (measured, not claimed)

Both Chinese and English text are scored. Chinese uses a character-based density model; English uses a word-based one. Verified against 6 control cases:

Draft	Score
English AI-cliche paragraph	74
The same paragraph, rewritten	24
Chinese AI-cliche paragraph	74
Human-written Chinese	0
Empty / very short input	refused, not scored

Inputs under 40 characters/words are refused rather than scored, because on a short snippet "no numbers, no first person" always fires and any score would be noise.

Honest limitations

Keyword matching is literal. If your rewrite quotes the cliches it is criticising — writing "eight of them pitched 'leverage'" — the detector still counts them and keeps a residual hit. It is a heuristic, not a judge. It also does not know your text's genre; a product spec will legitimately score as "no first person".

Who it's for

• Creators publishing to Chinese platforms who hit "low originality" flags
• Anyone whose editor or readers keep saying "this reads like AI"
• Agent users who want a scoring gate rather than a rewrite-by-feel prompt

What it is not

It is not an "AI detector bypass". It does not promise to defeat any specific detector, and it makes no claim about passing any platform's review. It makes the tells visible and fixes the writing. That's all it claims, and it does that reliably.

FAQ

Q: Does this guarantee my content passes AI detection? No. Nothing does, and any skill claiming that is lying to you. This skill identifies the specific structural tells that platforms and readers react to, and gives you six methods to fix them.

Q: Does the script need an API key or network access? No. Pure Python standard library, runs offline, reads a path you pass in.

Q: How do I run it? python scripts/ai_trace_check.py draft.md — or --text "...", or stdin via -, or --json.