Personal Voice

Personal Voice

Generic AI writing fails people twice: it doesn't sound like them, and everyone who knows them can tell. Style rules alone can't fix this because voice lives in measurable habits nobody thinks to state: how long sentences run, where the rhythm breaks, which words never appear. This skill extracts those habits from real samples, merges them with the rules the user CAN state, and holds every future draft to the result.

One boundary, stated once: this skill models the user's own voice at their request. Refuse to build a profile of another person from their writing in order to pass as them.

Mode selection

Check references/voice-profile.md in this skill's folder.

• Exists → application mode: read it and write to it.
• Missing → calibration mode: build it. Also enter calibration when the user asks to recalibrate or update their profile.

Calibration mode

1. Collect the inputs

Ask for 5 samples of the user's real writing, with this guidance, because sample quality caps profile quality:

• 150+ words each where possible; short fragments hide rhythm.
• Final writing the user was happy with, not rough drafts, and not text heavily edited by someone else (including AI): that would calibrate to the editor.
• Ideally matching the contexts they'll generate in. Five LinkedIn posts produce a LinkedIn voice; if they want emails too, ask for at least one email. If all samples are one genre, say plainly that the profile covers that genre and will extrapolate elsewhere.

Then ask for their explicit rules: preferred sentence lengths, rhythm notes, forbidden words and phrases, punctuation bans or loves, and the tones they want available (e.g., "dry, direct, occasionally warm; never peppy").

2. Analyze the samples

Measure, don't vibe. For each dimension below, extract the pattern AND its frequency, because voice is a set of rates, not a set of features. A signature move the user makes once per 500 words, deployed every paragraph, turns portrait into caricature — the most common failure of voice mimicry.

• Sentence mechanics: average length, range, fragment usage, how often a long sentence is followed by a very short one
• Paragraph shape: typical length, one-line paragraph usage
• Punctuation fingerprint: dashes, semicolons, parentheses, ellipses, exclamation points, oxford comma — presence and rate
• Vocabulary register: contraction rate, formality level, jargon comfort, profanity, favorite intensifiers
• Signature moves: how they open cold, how they close, how they emphasize (italics? repetition? short sentence?), humor style, hedging style
• Never-list: words and constructions absent from all samples that generic writing would have used (merge with the user's explicit forbidden list)

3. Reconcile rules vs. reality

Where explicit rules contradict the samples ("keep sentences short" but samples average 24 words), don't silently pick one. Ask: "Your samples run longer than your stated rule — match how you actually write, or shift toward how you want to write?" Aspirational voice is a legitimate choice; it just has to be a choice. Explicit rules win every tie the user doesn't resolve.

4. Write and install the profile

Fill in references/profile-template.md (bundled with this skill) and show the completed profile for approval. Include 2-3 short quoted snippets from their samples as touchstones. After approval, save it as references/voice-profile.md, keep the samples in references/samples/, and if a packaging tool is available, produce an updated .skill for the user to re-save so the profile travels to every future conversation. Without the re-save, calibration dies with the current chat — make sure the user knows this.

Application mode

1. Load the profile at the start of any writing task. Load the raw samples from references/samples/ only when the profile underdetermines something (a genre the profile is thin on, a long piece, a tricky tonal call) — the profile is the compression, the samples are ground truth.
2. Draft in the voice from the first word. Don't draft generically and then re-skin; structure and rhythm decisions happen at drafting time and can't be patched in afterward.
3. Flex by register within the voice. People sound like themselves differently in an email vs. an essay. Use the per-context notes in the profile; when writing in a context the profile doesn't cover, extrapolate conservatively and say you did.
4. Verify before delivering. Compare the draft against the profile numbers: sentence-length distribution in range? Any never-list word present? Signature moves at sample frequency, not caricature frequency? The test: would a colleague who reads the user's writing weekly pause on any sentence? Fix what fails, then deliver.
5. Learn from corrections. When the user edits your output or says "I'd never say that", that's calibration data arriving late. Append it to the profile's Learned Corrections section and offer to re-package. A correction captured once should never need repeating.

Precedence

When this skill is active alongside generic style guidance (including other skills), the user's voice wins conflicts. If the user genuinely writes with em dashes or loves a word that generic advice bans, their profile overrides the generic rule: the goal is sounding like them, not sounding like nobody.

Calibration failure modes

• Too few or too-short samples: proceed if the user insists, but mark the affected profile fields as low-confidence rather than presenting guesses as measurements.
• Samples that contradict each other: probably different registers, not noise. Profile them as separate contexts instead of averaging into a voice the user has never actually used.
• The user has no samples: offer rules-only mode — a profile built from explicit rules and a short interview. Honest but weaker; say so, and invite samples later.