Verified against Claude · 2026-07-25
Get Claude to write in your actual voice, not the generic AI default
A style-matching prompt that extracts a checkable checklist of concrete voice markers from several real writing samples before writing anything new, with a forced self-check pass to catch any drift back toward a more generic, more polished register.
The prompt
Ready to copy — highlighted parts are example details you can swap.
Below are 4 real samples of the voice I want matched. Before writing anything new, extract the pattern and show it back to me as a checklist of concrete, checkable markers — not adjectives like "friendly" or "professional," but things you could point to in the text: typical sentence length and rhythm, how paragraphs open, where contractions appear and where they don't, what gets a full sentence versus a fragment, how directly disagreement or bad news gets stated, any recurring words, transitions, or sentence structures, and how punctuation is used for emphasis. WRITING SAMPLES [Sample 1: a Slack update to the team] [Sample 2: an email to a client] [Sample 3: a paragraph from a blog post] [Sample 4: a reply to a customer complaint] The more varied these are in context and length, the more reliable the extraction — a single long sample risks the checklist capturing what's true of that one piece rather than what's true of the voice generally. After the checklist, wait for my confirmation before writing anything new — if I say the checklist missed something, revise the checklist first; do not just silently fold the correction into the new piece without updating the description I can check against. WRITE FOR THIS MEDIUM A LinkedIn post, roughly 120 words — adapt formatting conventions (line breaks, greeting or sign-off, structure) to what's normal for this medium, while keeping the checklist's actual sentence-level patterns intact; medium changes the container, not the voice. IF THE SAMPLES THEMSELVES DISAGREE If two samples pull in different directions on some pattern — one very short and clipped, another more expansive — say so in the checklist rather than averaging them into one blended rule that neither sample actually demonstrates. Note which sample the new piece will lean toward and why, based on A LinkedIn post, roughly 120 words and the brief, so the choice is visible rather than an invisible average nobody can check. ONCE THE CHECKLIST IS CONFIRMED, WRITE Announcing that we shipped a feature customers have been asking for since March. CONSTRAINTS - Match the samples' actual register, not an improved version of it. If the samples run a little informal, a little repetitive, or a little blunt, that is the voice — do not smooth it into more polished, more hedged, more evenly-paced prose just because it would read better in isolation. The single most common failure here is quietly reverting to a generic competent-AI register under the guise of "cleaning it up." - This one can run slightly more celebratory than the samples, since it's an announcement rather than a status update — but keep the same sentence rhythm and directness. - Avoid emoji even if LinkedIn posts in this space typically use them — none of the samples use emoji and that should hold here too. SELF-CHECK BEFORE YOU SHOW ME THE DRAFT Re-read what you wrote against your own checklist from earlier. Name any point in the draft where you drifted back toward a more generic, more polished register than the checklist describes, and either fix it or flag it explicitly — do not present a draft as a full match if you can see it isn't.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
Asking Claude to write 'in my voice' with no examples forces it to guess at a generic idea of a personal voice, which reliably regresses toward a smoothed, slightly-too-polished AI default; providing several real samples turns this into a few-shot pattern-matching task, a much better-supported capability than voice description from a one-line adjective. The forced intermediate step — extract the pattern into a checklist and wait for confirmation before writing anything — matters mechanically because it makes the extraction visible and correctable before any prose gets committed: if Claude describes the voice as warm and enthusiastic when what is actually distinctive is that sentences run long and hedge with 'I think,' that miss is visible in the checklist and can be corrected there, rather than discovered only in a finished draft that would otherwise need to be redone from scratch. Requiring the checklist itself to be revised on correction, rather than just folded silently into the next draft, keeps it as a durable, reusable spec for future pieces in the same voice instead of a one-time patch that has to be re-derived every time. The explicit instruction not to polish upward exists because Claude's default writing style is measurably more formal and more evenly hedged than most people's actual working voice, and left unconstrained it will quietly correct 'imperfections' in the sample voice — sentence fragments, blunt disagreement, a recurring pet phrase — treating them as things to smooth over rather than the actual signal being matched; naming this specific failure mode gives the model something concrete to check itself against, which matters because 'match the voice' alone is exactly the instruction that already failed to prevent this drift in an unconstrained version of the same ask. The closing self-check forces a second comparison pass against the model's own earlier extraction rather than a vague sense of whether it succeeded, which is the same self-referential mechanism that makes critique-then-revise prompting work generally, applied specifically to style fidelity.
Verified against
Claude Sonnet 4.6 · 2026-07-25
Changelog
- 2026-07-25 — Initial publish, verified against Claude Sonnet 4.6.
Building this for real?
This is a free starting point. If you'd rather have what Scult builds built and running for your business, that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
