Get the user's measured writing voice as explicit rules. CALL THIS BEFORE writing any prose the user will send or publish under their own name -- a post, email, reply, bio, doc, commit message, cover letter, or anything else in their name. The rules are measured from their own past writing (sentence rhythm, casing, punctuation rates, vocabulary, words to avoid), so they override your defaults. Cheap and fast; call it rather than guessing what the user sounds like.

Score a draft 0-100 for AI tells and get line-level findings. CALL THIS ON EVERY DRAFT before showing it to the user. Checks a curated slop lexicon plus structural signals -- sentence-length uniformity, paragraph shape, bullet symmetry, punctuation rates -- against THIS user's own measured baseline rather than universal rules. Above 85 reads human, below 50 is slop. Also returns the voice rules, so one call gives you both the critique and the target.

The impostor test: statistically, is this text a plausible sample of THIS user? Returns an authorship percentile against a null resampled from their own writing at the same length. Under 75 means indistinguishable from them; over 97 means it reads as someone else. This is a different question from voice_score -- that one catches AI tells, this one catches 'fluent, clean, and not you'. Run BOTH on any draft before showing it; a draft has to pass each. Also names the specific words that pushed it out of range.

Check whether a voiceprint exists and how solid it is: sample count, word count, when it was built, and whether the corpus is large enough for the measurements to be stable. Use this when you are not sure the user has set being-human up yet.

Build or refresh the voiceprint from the user's own writing. Sources: 'claude-history' and 'claude-projects' (local Claude Code history), 'chatgpt' (a conversations.json from a ChatGPT data export, needs path), 'files' (a folder of the user's own writing, needs path). Takes a few seconds. Run this once at setup, or again to fold in newer writing.

Record something about the user's voice that measurement cannot reach -- a phrase they have banned, who they are usually writing for, a correction they just made to your draft, a running joke. Use this whenever the user says 'I wouldn't say that', 'too formal', or rewrites something you wrote. Notes persist across every rebuild of the voiceprint and tend to matter more than the statistics.

Compile the voiceprint into a format another tool reads: 'chatgpt' (custom instructions box, ~1500 chars), 'agents' (AGENTS.md), 'cursor' (.cursorrules), 'claude' (CLAUDE.md), 'system' (raw system prompt), or 'json'. Use when the user wants their voice set up in a different assistant.