What problem does it solve?
Determining whether a piece of text was written by AI is usually answered by gut feeling or a single unreliable detector score. This Skill combines four independent measures — a heuristic pattern audit, deterministic statistical signals, an empirical Pangram model score, and a byte-level watermark scan — so the answer rests on evidence rather than vibes.
Core Features & Use Cases
- Heuristic Audit: Flags known AI writing tells (inflated vocabulary, recycled transitions, uniform rhythm) against a severity-tiered P0/P1/P2 pattern catalog, quoting each offending passage with no edits made.
- Statistical Signals: Runs a deterministic pass measuring n-gram entropy, type-token ratio, repetition, burstiness, and stylometry, each labeled with its research-backed reliability tier.
- Empirical Scoring: Submits text to the Pangram detection API for AI%/AI-assisted%/human% probabilities, with calibrated known-human baselines and batch draft comparison.
- Watermark Scan: Detects invisible characters, homoglyphs, bidi controls, and Unicode steganography embedded in the bytes, without any API key.
- Use Case: A writer asks which of three drafts reads most human. The Skill scores each draft plus two known-human passages in one batch, producing a ranked table where the relative gap — not any absolute number — drives the conclusion.
Quick Start
Ask the assistant to scan your document for AI writing patterns and score it for AI-detectability, for example: "Audit this blog post for AI tells and tell me how likely it is to be flagged as AI-generated."
