Brand Voice Guide

Extract how a brand actually sounds from what it has actually written, and turn it into a directory that a copywriter, a new hire, or a Claude agent can produce on-brand content from. One conviction drives it: a voice guide is extracted from real writing, not invented from adjectives. "We're bold, human, and authentic" describes nothing. A verbatim line the brand wrote, and a rule named from it, describes everything.

The output is a directory of reference files (not a single essay), so it can be consulted modularly at the moment of writing — the voice-core rules always, the relevant channel file for the piece at hand. It mirrors the structure a production content system consults before generating a word.

What this builds (and what it doesn't)

• It builds the voice directory: voice-core.md (the character), writing-standard.md (the concrete rules — a Do half and a severity-tagged Avoid half), per-channel files, real-examples.md, plus light about-company.md / icp.md context.
• It does not check drafts. Governing new content against the guide (AI tells, template language, brand-rule violations) is the job of a separate content-quality-check skill. This skill produces the standard that skill enforces — the Avoid section of writing-standard.md is the part it reads.

Inputs (harvest first, ask only for the gaps)

The skill is harvest-first: it extracts as much as possible from existing writing before asking the user anything. Collect what exists; note what's missing.

The single most useful ask: "Point me at the two or three things you've published that felt most like you — and one thing in your space that makes you cringe." The contrast defines the voice.

The workflow

Five phases. Harvest builds the draft; the interview only fills gaps; the Avoid rules come last. Show the user the tagged samples (end of Phase 2) and the draft files (Phase 4) before finalizing.

Phase 1 — Scope and intake

1. Confirm the brand one-liner and who the writing is for (reader/ICP).
2. List every channel the brand publishes on, and mark each intentional (a real voice, published with purpose) vs. passive (occasional, cross-posted without thought). Build channel files only for intentional ones. A brand whose Discord is its community needs voice-discord.md more than a twice-a-year Instagram needs voice-instagram.md.
3. Inventory the writing samples on hand; if there's little to no published content, switch to interview mode (Phase 3 does the discovery instead of the harvest).

Phase 2 — Harvest (extract voice from the samples)

Run the extraction passes in references/extraction-passes.md. In short: read everything and pull, verbatim, the language that could only come from this brand. Do not paraphrase — the specificity is the value. Capture:

• How they describe what they do (exact phrases), and any stated belief that could start an argument.
• The voice patterns that repeat: how pieces open, how they close, sentence length and rhythm, the relationship to the reader (explain / challenge / accompany), where the writing is most alive vs. most performed.
• Channel differences: how the register shifts between, say, email and LinkedIn.
• Recurring mechanics: preferred verbs and nouns, banned words, capitalization and number habits.

Output of this phase: a tagged bank of verbatim lines, grouped by the file each will feed. Show it.

Phase 3 — Interview to fill gaps (only what samples can't answer)

Some things rarely appear in polished content. Ask targeted questions from references/question-bank.md only where the harvest came up empty. The highest-value ones:

• The contrarian belief. "What does the brand believe that most competitors won't say out loud — the thing you'd say at dinner that might start an argument?"
• The relationship decision. "If the brand were a person at a dinner table, what role does it play?" Offer contrast pairs when they stall (mentor vs. peer; simplify vs. sit-in-complexity; projects-confidence vs. owns-the-uncertainty; challenges the reader vs. meets them where they are). The defense of the choice is more useful than the choice.
• Who it's NOT for, and what the brand would never say even if it worked for others.
• The one-sentence summary. "Write the sentence you'd want someone to say about the brand after three months of your content," then push once: "which word in that isn't quite true?"

In interview mode (no content to harvest), Phase 3 carries the whole discovery: work the question bank front to back, and lean on the AI-draft-reaction technique below to surface anti-patterns.

Phase 4 — Build the directory

Assemble each file using references/output-templates.md. Non-negotiable rules:

• Every file contains at least one real, verbatim sample the brand produced. A voice file with no real language is a horoscope. If you can't ground a file, it's a gap, not a guess.
• Every tonal pillar gets a named rule plus a verbatim do (and ideally a don't).
• Use the brand's exact words for rules ("we never sound like a professor who needs to impress you" goes in verbatim, not softened to "avoids academic tone").
• Build channel files one at a time; when a brand has 4+ channels, do the 2-3 highest-frequency ones first and add the rest in a second pass rather than producing shallow files for all of them.

Phase 5 — The Avoid rules (last) and verification

Build the Avoid section of writing-standard.md last, because an "avoid" only has meaning against a defined voice. It's the single home for banned words and voice-killing patterns (there is no separate anti-patterns file). Build it from three sources (detail in references/output-templates.md):

1. A universal starter list of generic/AI tells, which the user confirms or overrides.
2. The AI-draft-reaction: generate a short, styleless AI draft on one of the brand's own topics and have the user mark every line that makes them wince. These brand-specific tells are the most valuable ones and appear on no universal list.
3. The one industry phrase the brand refuses to use.

Tag each entry with severity (HIGH/MEDIUM/LOW) so the future content-quality-check skill can act on it. Then verify: read back the key decisions ("here's who you told me you are / who you're for / how you show up / what you never do — anything that doesn't sound exactly like you?"), and run the distinctiveness test on every voice file: if a line could be lifted onto a competitor's site unchanged, it isn't voice — rewrite or cut it. Close with a Gaps note (what's assumed vs. grounded).

Output

A directory, built with references/output-templates.md:

brand-voice/
├── about-company.md          # what it does, why it exists, the contrarian belief
├── icp.md                    # who it writes for (drop in icp-research output if you have it)
├── voice-core.md             # the character: tonal pillars (each with a verbatim sample), sentence
│                             #   rules, "persona we are NOT", the distinctiveness test
├── writing-standard.md       # the concrete rules — Do (words to use, formatting, caps, numbers) and
│                             #   Avoid (banned words + voice-killing patterns, severity-tagged; built last)
├── voice-<channel>.md        # one per INTENTIONAL channel: structure, hook patterns used/avoided,
│                             #   length targets, formatting
└── real-examples.md          # a small library of in-voice lines to reference for shape

Channel files are added only for channels the brand actively and intentionally publishes on. There is one home for "don't": the Avoid section of writing-standard.md (no separate anti-patterns file).

Composes with other skills

• icp-research — if it's been run, use its output as icp.md instead of re-gathering. Cross-link.
• content-quality-check — the enforcement side. This skill's voice-core.md and the Avoid section of writing-standard.md are what it checks drafts against.

Failure modes to avoid (the quality bar)

• Adjective soup — "bold, human, authentic." Describes nothing. Every rule needs a verbatim sample.
• Describing instead of extracting — writing what the voice should be rather than naming what it already is in the samples.
• Horoscope files — a voice file with no real brand language in it.
• Generic that survives the swap — a line that could sit on a competitor's site unchanged. Rewrite.
• Avoid rules written first — they're meaningless before the voice is defined; build them last.
• A file per dead channel — build only for channels published with intent.
• Paraphrasing the user's exact words — the specificity is the whole value; keep it verbatim.

Bundled references

• references/extraction-passes.md — the harvest engine: what to collect, and the passes that draft each file. Read at Phase 2.
• references/question-bank.md — the gap-filling questions, mapped to each file. Read at Phase 3.
• references/output-templates.md — the exact shape of every file, plus the universal Avoid starter list. Read at Phase 4/5.
• examples/ — worked examples built from real public copy. Each is a company folder holding the harvest (input.md) and the full multi-file voice directory the skill produced (brand-voice/): examples/cohere/ (primary; enterprise-sovereign, ownership-framed) and examples/notion/ (secondary; warm, playful-clear). Different voices, so the contrast shows the skill capturing what's actually distinctive.