The situation
Common Ground's partners put their names on drafts every day: emails, posts, proposals and bids. Every one of those drafts needs to sound like the person whose name is on it, not like an assistant being careful. The tools most teams reach for to do that, a paragraph of adjectives ("warm, direct, witty") or a few pasted examples in a prompt, do not hold up. Research on frontier models imitating real authors found that in-context examples alone still drift toward a generic, averaged tone, and past four or five examples, more barely helps.
We built a different pipeline: a counted fingerprint of how a person actually writes (function-word frequency, sentence length, punctuation and casing habits, each measured against a plain AI baseline), a bank of real passages tagged by register and retrieved a few at a time rather than pasted wholesale, a set of contrastive rules learned from what the person corrects when a machine cleans up their words, and a scorer that checks a draft's stylistic distance from the real thing. The whole system is judged by one test: can the person tell their own writing from a generated draft on the same task. Failing to tell them apart is the bar, not a compliment.
A style guide built from adjectives and a handful of pasted examples produces writing that reads as generically competent and is easy to spot as AI-drafted, because style lives in small, countable choices (function words, sentence rhythm, punctuation) that an adjective list does not capture and a few examples do not transfer.
Built a counted stylistic fingerprint per register from a person's own real writing (texts, email going back years, dictation, spoken transcripts), an exemplar bank retrieved a few passages at a time rather than pasted wholesale, a contrastive rule layer learned from the edits a person makes to a machine's drafts, a scorer, and a blind test where the person tries to tell their own writing from a generated draft.
The first run read one partner's texts, email, dictation and call transcripts. The first version of the scorer was provably broken (it called the majority of the person's own emails AI-written) before it was fixed and proven on real text. The output now feeds every drafting surface that carries a partner's name.
Read on
The decision and the reasoning
Why we did it this way, told first.
ReadWhat we did and what it produced
The work, decision by decision, and the result.
ReadA slice of the project list
A few related projects.
ReadPrivate side
The full report, the source files on record, the timeline and every figure with its source. Password.
Enter