Agent Skills: Analyze Voice

Analyze Voice

A voice profile is what stops the team's copy reading like generic AI. This skill turns a writer's real samples into Voice.md — the file every writer and editor reads before they touch a line. Your job is to find the patterns that make the samples sound like one person, not to invent a voice.

Decide the tier by sample count

The honest output depends on how much real writing you have. Pick the tier, then build to its confidence ceiling — never claim more.

• 0 samples → aspirational. No writing to analyze. Run a guided questionnaire (below) and write a target voice from the answers. Mark every section aspirational in the metadata and add the line: unproven — written from intent, not samples. Regenerate from real writing before trusting it.
• 1–4 samples → provisional. Enough to sketch, not to confirm. Capture what the samples show, mark thin sections provisional, and note the profile firms up at 5+ samples.
• 5+ samples → full. Reverse-engineer across all samples. Keep only patterns that recur in most of them; drop one-off quirks. This is the only tier that earns full confidence on its core sections.

Capture by reverse-engineering

For each sample, ask: what would I have to tell a writer to produce exactly this? Every answer is a captured rule — a word choice, a sentence shape, a punctuation habit, a thing the writer refuses to do.

• Across 5+ samples, keep the rules that recur in most samples and discard the rest. Recurrence is the test — a pattern in one sample is a coincidence, a pattern in four is the voice.
• Pull signature phrases verbatim. Exact recurring openers, transitions, and sign-offs are the highest-fidelity part of a voice. Copy them word for word; never paraphrase.
• Separate the structural from the personal. Product-voice patterns (register, banned words, sentence length, how it handles a feature) reproduce reliably. Subtle personal rhythm does not — see the ceiling below.

Capture with no samples (questionnaire)

When there's nothing to analyze, ask the writer — one question at a time:

• Name three writers, brands, or pages whose voice you'd want to sound like, and one you'd never want to sound like.
• What does your product never say? (slang, hype words, jargon, emoji, exclamation marks)
• Formal or casual? Do you address the reader as "you"?
• Write one sentence describing your product the way you'd actually say it out loud.

Build the aspirational profile from the answers. It is a target to test real writing against, not a description of writing that exists.

Validate before you ship the profile

Run these checks on the drafted Voice.md — fix what fails:

• is/is-not presence — every entry in the identity core pairs a concrete "is" with a concrete "is-not". A lone "is" ("clear and friendly") describes nothing; the contrast ("friendly, not chummy — warm without slang") is what's usable.
• terminology presence — the use/avoid term lists are non-empty and drawn from the samples, not generic. Generic banned-word lists belong in the editing skill; this section captures this writer's specific words.
• reproduction test — write one line in the captured voice and read it against the samples. If it doesn't pass as the same writer, the fingerprint is too thin — capture more, or drop the tier.

The honest ceiling

Structured product voice reproduces well; a subtle personal voice does not. Say so in the profile's metadata. Capturing register, vocabulary, and refusals gets the team most of the way; the last increment of a distinctive human cadence won't survive the round-trip. Over-claiming fidelity is the failure mode this caveat exists to prevent — a writer who trusts an over-confident profile ships copy that sounds almost-right, which is worse than obviously-generic.

Write the profile

Write Voice.md into the project's working directory — the project the copy is for, never inside this profile or its repo (that holds instructions, not work products). Normal case — never VOICE.md. Follow the section-by-section schema in voice-schema.md; the schema is the single owner of the file's structure, so build every section it lists, at the confidence the tier allows.