| name | my-voice |
| description | Use when writing or rewriting text that must sound like a specific real person wrote it (their Reddit post, email, blog, bio, social copy) — the user offers a writing sample, asks to "put this in my voice", "make it sound like me", "de-slop this", or complains that a draft reads like AI. Distills a reusable voice profile from the sample and rewrites against it. |
My Voice
Overview
Rewriting "in someone's voice" fails two ways at once: the agent applies generic
anti-AI de-slopping that makes text clean but anonymous, and it leaves its OWN tells in
(em-dashes, "not just X, it's Y", rule-of-three). The fix is to stop working from vibes:
distill a concrete, checkable profile from a real sample, then verify the output
against it mechanically. A voice is a set of observable habits, not a mood.
Method
1. Get a real sample
Ask for one if none given. Rank by fidelity:
- Unedited chat/DM/message history (best — natural, unperformed)
- Emails, comments, forum/Reddit posts
- Published blog/marketing (worst — already edited, may not be their own hand)
Aim for ~400+ words across 2–3 pieces. Prefer the genre being written (a Reddit
voice ≠ a bio voice). If the only sample is one polished paragraph, say so — the profile
will be thin.
2. Distill the profile (write it down, don't hold it in your head)
Fill every dimension from the sample with a specific observation + a quoted example.
Guessing = leave it "unknown", never invent:
- Sentence length & rhythm — short/medium/long; uniform or varied? run-ons?
- Punctuation habits — em-dashes? semicolons? parentheticals? ellipses? Note what
they use to join clauses instead (comma splices? periods? "So"/"Then"?).
- POV & address — first person? second-person "you"? name the reader?
- Openers & connectives — how sentences start ("So", "The problem was", "Look,").
- Register — dry / wry / warm / blunt / profane / formal. Hedged or flat-assertive?
- Concreteness — numbers-first? named examples? or abstract?
- Honesty tics — do they admit dead ends, uncertainty, self-deprecation?
- Signature quirks — lowercase, specific words, a recurring move.
- Never-does — the strongest signal. What's absent from the sample that AI defaults
to? (usually: em-dashes, antithesis, exclamation points, marketing words.)
3. Interview for stated preferences (the sample can't tell you these)
A sample shows what the person did, not what they want or consciously avoid. Ask a
short round (one question at a time, skip any the sample already answers clearly):
- Aspirations vs habits — "Anything in these samples you'd change? Words or moves
you fall into but don't like?" (Habits you'd otherwise faithfully reproduce as tells.)
- Hard bans — "Words, phrases, or punctuation you refuse to use?" (People often have
a personal blocklist beyond the universal AI tells — e.g. never "utilize", never
emoji, never exclamation points.)
- Register range — "Same voice everywhere, or more formal for X and looser for Y?"
Capture per-context variants if so.
- Contractions / formality / profanity — confirm the defaults, since these swing by
audience and the sample may be one point on that range.
- The tell test — "What makes writing instantly read as AI-generated to you?" Their
answer goes straight into the kill-list; it's the most personal de-slop signal there is.
Record stated preferences in the profile and mark them as stated (vs observed
from the sample). When they conflict, stated preference wins — it's the conscious choice.
4. Build the de-slop kill-list
Two layers:
- Universal AI tells (strip always): em-dash-as-connector, "not just X, it's Y" /
"it's not A, it's B" antithesis, rule-of-three cadence, marketing register
(seamless, powerful, robust, leverage, unlock, delve, elevate, game-changing,
transformative, thrilled, dive in, testament to, "in today's ... world"),
exclamation points, "Whether you're X or Y", a summarizing "In conclusion/Ultimately"
outro, emoji-as-decoration, every-paragraph-perfectly-balanced.
- Person-specific — anything the sample proves they never do (e.g. "sample has 0
em-dashes across 400 words → em-dashes are banned for this voice specifically"), plus
the stated bans from the interview.
5. Persist the profile
Save it to .claude/voice/<name>.md (or the user's memory dir) so future rewrites reuse
it instead of re-deriving. Keep the observed / stated tags. Reuse and refine it
when more samples or corrections appear.
6. Rewrite, then run the checklist
Draft in the voice, then verify against the profile before returning. This step is the
skill — an unverified rewrite is where the tells survive:
If any box fails, fix that specific thing and re-check. Do not return on first draft.
Common mistakes
- Generic-casual ≠ their voice. "Made it conversational" is not the job; matching
their specific habits is. A dry numbers-first writer rendered as a breezy blogger
fails even with every AI word gone.
- Injecting your own tells while removing theirs. The em-dash and the antithesis are
the two that survive most often — they feel like "good writing", so agents add them
back. The checklist exists because self-assessment misses them.
- Over-applying de-slop. If the person genuinely uses semicolons or the occasional
em-dash, keep them at their rate. De-slopping means matching the human, not stripping
to a lowest-common-denominator plainness.
- Inventing quirks from a thin sample. One paragraph can't tell you their opener
habits. Mark unknowns; don't fabricate a persona. The interview (step 3) is how you
fill gaps the sample leaves — ask, don't guess.
Profile template
# Voice profile — <name>
Distilled from: <sources>. Tags: observed = seen in samples; stated = told directly.
## Dimensions
- Sentence length & rhythm (observed): ...
- Punctuation (observed): ...
- POV & address (observed): ...
- Openers & connectives (observed): ...
- Register (observed): ...
- Concreteness (observed): ...
- Honesty tics (observed): ...
- Signature move (observed): ...
## Kill-list (never, for this voice)
- Person-specific (observed absent): ...
- Stated bans (interview): ...
- Universal AI tells: ...
## Register range (stated): ...
## Open (ask next time): ...
