How I Finally Got AI to Write Like Me (Mike Taylor, Every)
Why this is in the vault
A concrete, tested method for building a personal writing skill: held-out posts, substance-only briefs, a pairwise judge, and an optimizer. It is the first evidence-backed alternative to "have Claude write a style guide from my posts".
The core argument
Asking Claude to write a style guide from your past posts yields a caricature of your greatest hits. In Taylor's tests that Claude-written skill lost to all three optimizers on 80-95% of held-out posts. His six-step process:
- Gather 12-25 published pieces from one format (or meeting transcripts); hold some out as a final exam; skip tweets and emails as source material.
- Turn each post into a brief carrying thesis, facts, and anecdotes but none of the style, so a draft from the brief can be attributed to the skill. He frames a "distillation ratio": input tokens should outnumber output tokens.
- Build a pairwise "gold triad" judge: it sees the real post plus two drafts, picks the closer one, runs twice with order swapped, and writes a note on how each draft differs. Opus 5.5 was consistent 86% of the time versus 70% for Sonnet 5.5.
- Optimize the skill instead of authoring it (autoresearch, DSPy GEPA, ce-optimize). GEPA won head to head; most gains came within three rounds; cost was roughly $27-53 per skill versus about $5 one-shot. Optimized skills are made of corrections drawn from judge notes ("no invented jokes", no "Here's the thing"), not style descriptions.
- Optimize per writer model: Opus 5.5 went from 16% with no skill to 93% with its GEPA skill, and an optimized Sonnet 5.5 skill beat Opus with the Claude-written one on 46 of 60 judgments.
- Write from a rich brief (not a one-liner), then feed your edits back as training data. Caveat: the judge rewards resembling the writer you were, so hold out your most recent posts if your style is moving.
Takeaway: define the eval before optimizing the task. The skill does not decide what is worth saying.
Mapping against Ray Data Co
Directly bears on the writing-standard contract (WRITING-rdco.md) and the Sanity Check voice work: our voice rules are currently a hand-authored description, which is exactly the "informed guess" Taylor shows loses to optimized skills. Concrete adoption path: build a held-out eval from published Sanity Check and Ship 30 posts (founder-written only, per the voice-from-real-signal rule, so unpublished drafts do not count), use a swap-order pairwise judge on Opus, and let GEPA iterate the voice skill from judge notes rather than from our rule list. His "no tweets/emails as corpus" finding also supports keeping X voice separate from long-form. His distillation-ratio idea is a usable gate for founder-approved ghostwriting: Ray drafts only from a substance brief the founder supplied. Caveats: single author, self-reported metrics, N=30 held-out; and his judge test showed Jev unreliable for creative judgment, which tempers our "Jev for typed sets" habit for style scoring (it remains fine for classification). Not decision-forcing; a candidate experiment once a published-post corpus of 12+ pieces exists.
Related
- [[2026-03-12-every-ai-writing-style-science]]
- [[2026-03-19-every-ai-style-guide]]
- [[2026-09-09-every-writers-who-use-ai]]
- [[2026-07-31-every-voice-ai-workflow]]
- [[feedback_voice_from_real_signal_only]]
- [[feedback_writing_standard_contract]]
- [[feedback_jev_for_typed_set_problems]]
⚠️ Sponsorship
Mid-issue paid placement for Atlassian's Jira Product Discovery (roadmap/prioritization pitch); one-off per Every sponsor pattern, unrelated to the article. House CTAs also present: Every All Access subscription with "Builder Pack", Every Agent, Monologue, and the author's own skill on GitHub that "will soon be part of Every's Compound Writing plugin" (self-promotion; ce-optimize is an Every Compound Engineering tool and was one of the winning optimizers, so Every-product favoritism is possible). Author is Every's head of evals.