voice-define

/voice-define

Brand voice capture. Use this skill the first time when bootstrap-marketing walks through it, and every few months as new content ships and the voice should be re-grounded against real artifacts instead of synthesized hopes.

Voice transfers better from samples than from rules. The output of this skill is marketing/VOICE.md — a file that copywriter and brand-reviewer read on every campaign.

Three paths (mix freely)

Path A — From scratch

For brand-new projects with nothing shipped. Walk through 5–8 questions covering posture, reading level, formality, sentence rhythm, emoji policy, words-you-love, words-you-hate (→ do_not_say), and brands admired (→ Path C anchors).

Output is rules-based, no examples — the weakest of the three paths in isolation, but a reasonable starting point that gets upgraded later.

Path B — From existing examples

For projects with shipped content. Ask for 3–5 pieces (paths or URLs) that exemplify the desired voice. Read them, then extract:

• Cadence — average sentence length, range, variance
• Vocabulary distribution — formal / casual / technical / colloquial
• Recurring stylistic moves — em-dashes, parentheticals, lists, asides, footnotes, all-caps emphasis, italics
• Hook patterns — how openings work (question / scene / claim / contrarian / code-first / narrative / hot-take)
• CTA patterns — how closings work
• House vocabulary — words that recur naturally (good candidates for positive voice notes)
• Absent words — words conspicuously avoided (candidates for do_not_say)
• Structural patterns — how each piece is organized, where the payoff lands

Write the extraction with citations back to the source pieces. Future copywriter runs read the exemplars, not just the rules.

Path C — Anchored to reference brands

For projects that want to sound like a specific other brand (or a blend — blends are usually more interesting than single anchors).

Ask: "Name 2–4 brands whose voice you want to emulate. URLs to specific pieces of theirs help a lot." Examples:

• B2B / SaaS: Linear, Vercel, Stripe, Anthropic
• Indie / founder voice: 37signals, Pieter Levels, Patrick McKenzie, Paul Graham
• Newsletter / longform: Ben Thompson, Matt Levine, Packy McCormick
• Developer-focused: Cloudflare blog, Fly.io blog, Tigris blog

For each named brand:

1. If a URL is given, fetch and read at least 2 pieces — one long-form, one short. Extract the same patterns Path B extracts.
2. If no URL, summarize the brand's known voice from training context and flag it explicitly in VOICE.md as "anchor from memory, not from fresh sample." If accuracy matters, ask for URLs.

Then write the blend, not the union — describe how the brand combines anchors:

Aspirational blend: Linear's clarity (no jargon, no warm-up) + Stripe's specificity (numbers in headlines, concrete examples) + 37signals' contrarian edge (state the unpopular position, defend it).

Cite the original sources so future runs can re-check the synthesis.

Mixing paths

Common combinations:

• A + C — brand-new project, no content yet, but knows what it wants to sound like. Path A nails the personality questions, Path C provides concrete style anchors.
• B alone — project has shipped enough content that the voice is already coherent; just needs codifying.
• B + C — project has shipped content and knows what it wants to evolve toward. Path B captures where the voice is; Path C captures where it should go. VOICE.md calls out both: "Current voice (from shipped work):" and "Direction (from anchors):".
• A + B + C — full mix, best for thorough first-time capture.

Always do first

Read .claude/project.yaml and confirm there's a marketing: block. If not, surface — run /bootstrap-marketing first.

Check if marketing/VOICE.md already exists:

• Doesn't exist → fresh capture; ask which path(s)
• Exists → confirm whether this is a refresh (re-grounding against new shipped content), an addition (adding a new path's output), or a full rewrite

Output structure

Write to marketing/VOICE.md using templates/VOICE.md.tmpl. Sections:

# Brand voice — <project name>

Captured: <date> · Paths used: <A / B / C / mixed>

## Posture

<one-paragraph distillation of the brand's stance on the world>

## Cadence

<sentence-length stats from Path B, or declared rhythm from Path A>

## Vocabulary

### House words (use freely)

- <word>: <why it fits>

### Aspirational anchors (Path C)

- <brand>: <what we're borrowing from them>, sources: [piece 1](url), [piece 2](url)

### do_not_say (mirrored into manifest)

- <word>: <why it's banned>

## Stylistic moves

- <move>: <when to use it>, example from <source>: <quoted line>

## Hook patterns

<per-channel hook conventions, with examples>

## CTA patterns

<how closings work, with examples>

## Positive exemplars (path B + C citations)

- [<title>](<url or path>) — <why this is canonical>

## Negative exemplars (what we're NOT)

- <description> — example we explicitly don't want to sound like

## Notes for the copywriter

<plain-English notes that an agent reads at runtime — "open with code,
not with story" or "the second paragraph should always have a number">

What this skill mutates

• marketing/VOICE.md — the voice file itself
• .claude/project.yaml — only marketing.brand.do_not_say and marketing.brand.house_examples (the file pointer list); never touches unrelated manifest fields

What this skill doesn't do

• Decide whether the voice is "right" — that's a judgment call only the human can make. Print the synthesis, ask "does this feel like you?", iterate if no
• Modify already-published content
• Run against private examples without permission to read them (URLs behind auth need the human to paste the content)

Re-run cadence

Suggest re-running this skill:

• After every 5–10 shipped campaigns (Path B re-grounding)
• When the brand consciously evolves (new positioning, new product)
• When brand-reviewer's quarantine list grows — the voice file may be out of date with what's actually working