redtone

What it does

“Make it sound human” is the wrong ask — humanity in general is an average human, which is nobody again. This skill does something else: it matches text to a SPECIFIC author's hand and verifies the fit with a number. The profile is derived from your own chats and written pieces: smiley-paren frequency and run length, dashes per 1000 characters, sentence-length variance, words per paragraph, signature particles and — the strongest signal — words the author physically never uses. Thresholds are not invented but derived by negative control: a rule that flags more than 5% of the author's authentic texts is wrong by construction. The method's key finding: one person's registers are different dialects, not shades. A chat-derived profile rejected 4 of the same author's 4 written texts; a synonym preference flipped from 15:1 to 0:13, and lowercase sentence starts went from “almost always” to “never”.

How to use

1. 1Collect two corpora: a chat export (2–3 chats across registers, from ~1500 of your own messages) and separately 3–5 written pieces — posts, letters, articles. They are not interchangeable: chat gives punctuation and lexicon, writing gives rhythm.
2. 2Measure what you don't know about yourself. Asking the author is useless: in our run the person was sure they often used ellipses and slashes — the corpus showed a statistical zero. People remember intent, not habit.
3. 3Find negative markers — words appearing 0–2 times per hundred thousand. “Never write these six words” beats “write more vividly”. Check the list for homonyms and fixed phrases: ours flagged the Russian word for “data” as bureaucratese.
4. 4Separate registers and let them override base rules ENTIRELY, forbidden-word list included. A word banned in chat may be normal in writing — that's a measurement, not an inconsistency.
5. 5Calibrate by negative control: apply the rule to the author's authentic texts, count the rejects, ceiling 5%. Our “common sense” rules rejected 12%, 8% and 7.7% — all had to be rewritten.
6. 6Hold out material and never calibrate on it. Re-run every rule change against it: ours confirmed rhythm transfer to unseen text and simultaneously exposed two defects no training text revealed.
7. 7Add a positive control: a typical AI text with zero forbidden words must still be caught by rhythm. A profile that flags nothing is useless.
8. 8Fix the tool contract: 0 clean, 1 divergences, ≥64 could not measure. A missing result must never read as “nothing wrong”.

When it triggers

Example

The check is necessary but not sufficient: our first draft passed with zero divergences — and contained an em dash inside the sentence “I never use em dashes”. The threshold was lenient and the metric honestly stayed quiet. Numbers catch systematic drift from your hand, not silliness — you still have to read it yourself.