Govern the BRAND-VOICE.md — extract a brand's writing voice from URL, Notion page, MD file/directory, or interview into a structured executable doc; update incrementally; diff versions; validate the canonical format; show testable rules. Supports multi-voice via `voice.extends` (founder on corporate, persona on institutional, multi-host) with `_replace`/`_remove` overrides. Consumed by writing skills via `-f` (e.g. `humanize-en`). Use when defining, extracting, refreshing, validating, or inspecting a brand's writing voice — "create a brand voice doc", "extract voice from this site", "tone of voice", "writing style guide", "BRAND-VOICE.md", "founder voice", "persona voice", "multi-voice".
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Govern the BRAND-VOICE.md — extract a brand's writing voice from URL, Notion page, MD file/directory, or interview into a structured executable doc; update incrementally; diff versions; validate the canonical format; show testable rules. Supports multi-voice via `voice.extends` (founder on corporate, persona on institutional, multi-host) with `_replace`/`_remove` overrides. Consumed by writing skills via `-f` (e.g. `humanize-en`). Use when defining, extracting, refreshing, validating, or inspecting a brand's writing voice — "create a brand voice doc", "extract voice from this site", "tone of voice", "writing style guide", "BRAND-VOICE.md", "founder voice", "persona voice", "multi-voice".
when_to_use
Routes via `$ARGUMENTS` first token — `extract` (sources → BRAND-VOICE.md; `--extends <parent>` scaffolds a child), `update` (refresh from new sources), `diff` (regression check; single-arg form when child has `voice.extends`), `validate` / `lint` / `check` (walks chain), `show` (testable rules; `--chain`/`--explain`/`--raw` for inheritance). Skip when the user wants to apply or check prose against an existing voice (rewrite, humanize, "does this match") — invoke `/humanize-en -f BRAND-VOICE.md` instead.
These rules govern every prose artifact this skill emits — READMEs, CHANGELOGs, commit messages, PR bodies, release notes, doc paragraphs, non-trivial comments. Apply them at draft time, verify before output.
Match the surrounding style — punctuation, capitalization, backtick conventions, em-dash vs parens, bullet style.
Every sentence changes the reader's understanding. Cut it otherwise.
Front-load the verb — "Creates", not "This helps you create".
Concrete over abstract. Lists for ≥3 enumerable items.
Assert positively. Reserve negation for real constraints (NEVER commit secrets).
No marketing words: powerful, robust, seamlessly, leverage, unlock, comprehensive, delightful.
No AI tells: delve, tapestry, intricate, pivotal, testament, underscore, crucial, garner, showcase, additionally, moreover, furthermore, indeed.
After drafting English prose, invoke /humanize-en if installed.
Govern BRAND-VOICE.md — the canonical writing voice document for a brand. Two layers: YAML frontmatter (machine-readable normative rules consumed by writing skills) plus eleven prose sections (human-readable rationale). Same split as DESIGN.md and the same design-system skill pattern: a canonical file at the project root, CLI-style subcommands for the lifecycle.
Additional context from the user: $ARGUMENTS
Subcommand routing
Parse the first positional token of $ARGUMENTS. If it matches a verb below, load the referenced file and follow its workflow. Otherwise fall through to the default workflow at the end of this document.
First token
Mode
Reference
extract
Ingest sources, synthesise canonical voice doc, write to ./BRAND-VOICE.md
There is no apply subcommand. Application of the voice — rewriting prose to match it — is the consumer skill's job. The current consumer is humanize-en -f BRAND-VOICE.md, which calls scripts/extract_rules.py --full for chain-resolved flat rules. Any other skill that needs the voice contract follows the same path; alternative consumers can read the YAML frontmatter directly when their allowed-tools excludes Bash.
Canonical file location
The voice doc lives at ./BRAND-VOICE.md by default — at the project root, versioned in git, alongside DESIGN.md, README.md, LICENSE.md. Override the path with -o <path> when the voice is multi-project.
extract refuses to overwrite an existing file. To refresh, use update. To replace, delete first.
When -s is passed alongside extract, the skill also writes a copy to ~/.agents/output/{project}/brand-voice/brand-voice-{slug}.md for pipeline-history consumers ({slug} = kebab of voice.name; {project} = kebab-cased basename of the git toplevel, else cwd) and reports its fully-expanded absolute path (no tilde, no magic). The canonical file at ./BRAND-VOICE.md remains the single source of truth.
Cross-repo distribution
When a brand spans multiple repositories (a www repo, an iOS repo, a docs repo, a marketing site), the same BRAND-VOICE.md should govern all of them. Pick one pattern and document it in each project's CLAUDE.md:
Brand workspace canonical — keep BRAND-VOICE.md at the brand workspace root (e.g. ~/<brand>/BRAND-VOICE.md). Each subproject references it via absolute path: /humanize-en -f ~/<brand>/BRAND-VOICE.md draft.md. Simplest. Best when subprojects share a local workspace.
Monorepo — packages/brand/BRAND-VOICE.md consumed by every app in the monorepo. Single PR for cross-cutting voice changes.
Git submodule — canonical brand repo included as a submodule. Atomic updates via submodule bump. Best when the brand is owned by a separate team.
Published package — @<org>/brand-voice on npm with BRAND-VOICE.md plus the bundled scripts (extract_rules.py, voice_lint.py). Versioned, works cross-repo without a shared filesystem.
Copy + periodic /brand-voice diff — a copy in each repo; periodic diff <canonical> <local> catches drift. Simplest tooling, highest drift risk. Pair with a CI check.
Notion-as-source-of-truth is its own pattern: keep the spec in Notion, refresh local BRAND-VOICE.md periodically via /brand-voice update -n <page-id>. Notion stays the editorial surface; the local file is the executable artifact.
Multi-target: one file or many?
The default and recommended pattern is one BRAND-VOICE.md per brand. Within that file, contexts: handles register variation across document types (RFC vs landing page vs press release), audience segments (B2B vs consumer, technical vs lay), or channels (long-form vs social vs email):
Different contexts share the same lexicon, the same forbidden patterns, the same pronouns — what changes is the register, the sentence rhythm, the example openers.
Multiple voice files are warranted only when the brand has genuinely separate sub-brands with separate voices: a luxury group that owns Maison X Couture (institutional, French-rooted) and Maison X Beauty (more accessible, broader audience). Each sub-brand gets its own BRAND-VOICE.md. The skills consume each independently — humanize-en -f maison-x-couture.md for one, humanize-en -f maison-x-beauty.md for the other.
Inheritance via voice.extends — when sub-voices share a common substrate (founder voice on top of corporate, persona on top of institutional, multi-host media brand), declare voice.extends: ./BRAND-VOICE.md on the child file. The child inherits the parent's rules and overrides only what differs. Per-field merge policy, _replace / _remove overrides, cycle detection, and validation order live in references/canonical-format.md § Inheritance; a worked example sits in references/example-multi-voice.md.
When in doubt, start with one file. Adding contexts.foo later is cheaper than splitting two files later. Adding voice.extends later, when a real second voice emerges, is cheaper than over-engineering inheritance up front.
Source resolution
Sources are combinable — pass any number of -u, -n, -d, -f. The skill aggregates all sources into a working draft, then synthesises the canonical format once.
Flag
Source
Mechanism
-u <url>
URL
WebFetch (or your harness's web-fetch tool) direct → fallback /markitdown -s <url> if binary/error
-n <id|url>
Notion page
Notion MCP fetch tool (page + linked sub-pages, depth 1) — no MCP: export to Markdown and use -d
-d <dir>
Folder of MD/MDX
Glob <dir>/**/*.md → aggregate
-f <file>
Single MD/MDX/TXT
Read direct
(none, with extract)
Interview
8 canonical questions via AskUserQuestion (when unavailable, ask in plain text)
Full resolution rules — including failure handling, conflicts, MCP unavailability, large-folder fan-out, and contribution summary — live in references/source-resolution.md.
The Notion MCP is authorised through Claude Code's permission layer, not via this skill's allowed-tools. If the MCP is not installed, -n errors with a clear install pointer and the workaround (export Notion → MD, then -d).
The split is deliberate. Tooling reads YAML; humans read prose. Consumers like humanize-en -f BRAND-VOICE.md load only the rule block (50–150 lines via extract_rules.py --full), not the full doc, so the voice doc can be richly explained without bloating downstream contexts.
Pipeline integration
Brand voice is consumed by writing skills via -f. The current consumer is humanize-en:
/brand-voice extract -u https://example.com/about
→ ./BRAND-VOICE.md
/humanize-en -f ./BRAND-VOICE.md draft.md
→ draft humanized against universal AI tells + brand-specific rules
$SKILL_DIR = this skill's folder — ${CLAUDE_SKILL_DIR} in Claude Code, the directory containing this SKILL.md elsewhere.
Two ways for a consumer to read the rules:
Invoke extract_rules.py --full — preferred. The script flattens the YAML to plain text, automatically resolves any voice.extends chain, applies _replace and _remove overrides, and emits a 50–150 line block ready for inclusion in an LLM prompt. This is what humanize-en -f does as of the inheritance release.
Read the YAML frontmatter directly — fallback when the consumer's allowed-tools does not include Bash, or when the consumer wants raw structure. The consumer parses the YAML and uses forbidden_lexicon, rewrite_rules, sentence_norms, forbidden_patterns, pronouns, core_attributes, contexts directly. This path does not resolve voice.extends — child files appear as-written.
Both shapes are documented in references/schemas.md § extract_rules.py. The --legacy flag emits the v1 minimal output (byte-identical to the pre-inheritance shape) for any external consumer pinned to it.
When a brand voice rule conflicts with a universal AI-tell pattern (e.g., the voice requires em-dashes vs pattern #14), the brand rule wins — it is the user's contract. Conflicts are logged in the consumer's report.
Validation — voice_lint.py
Every doc the skill writes (or the user authors) is validated by scripts/voice_lint.py:
Verdicts: GREEN (zero errors, zero warnings), YELLOW (warnings only — acceptable but flagged), RED (errors — block). Output is JSON per references/schemas.md § voice_lint.py.
extract and update lint before writing. RED → fix and re-lint without prompting the user. YELLOW → present warnings to the user and proceed on confirmation.
Default workflow (no subcommand)
When the first token of $ARGUMENTS does not match extract|update|diff|validate|lint|check|show, the skill behaves as follows:
No BRAND-VOICE.md at the target → suggest /brand-voice extract with the sources the user mentions inline. Do not silently extract.
BRAND-VOICE.md exists → run show --rules and print the testable rules. Useful when the user types /brand-voice to glance at the current contract.
The argument looks like a URL → suggest /brand-voice extract -u <url> (or /brand-voice update -u <url> if a doc exists).
The argument looks like a file path → suggest the corresponding -f invocation.
The default workflow exists to avoid silent state-modifying actions. Every write goes through an explicit subcommand.
Rules
Canonical file is git-versioned. Treat ./BRAND-VOICE.md as a code asset. Diff before merge. The git history is the audit trail.
Lint before write. Every extract and update runs voice_lint.py on the synthesised content before the user is asked to approve. RED never reaches disk.
Never overwrite.extract refuses to overwrite an existing file. update always shows a diff and asks for explicit yes. diff is read-only by definition.
Manual sections are sacred. A section marked <!-- manual: true --> is preserved verbatim by update. Do not re-synthesise.
Conflicts surface, never override. When two sources disagree, surface via AskUserQuestion. When a source contradicts the existing doc during update, surface. No silent picks.
Output paths follow the repo contract. Default canonical at ./BRAND-VOICE.md. Pipeline copies under ~/.agents/output/{project}/brand-voice/brand-voice-{slug}.md only when -s is passed.
When to defer to another skill
Apply the voice on a prose draft → /humanize-en -f BRAND-VOICE.md <draft>. This skill never humanises.
Convert a non-Markdown source to MD first → /markitdown -s <source>, then /brand-voice extract -f <markitdown-output>.
Extract design tokens, not voice → /award-design and /design-system. Brand voice is prose; brand visuals are tokens. Different docs, different lifecycles.
references/canonical-format.md — full schema, required vs recommended sections, section ordering, manual-section markers, inheritance via voice.extends. The contract.
references/example-chanel.md — complete reference voice doc, anchored on chanel.com primary sources (Métiers d'art savoir-faire page, House of Chanel history, founder page) plus Met Museum and Wikipedia as cross-references. Use as a structural template.
references/example-multi-voice.md — worked example of voice.extends: a fictional founder-led startup with parent + child + merged result side-by-side, plus when to use _replace vs _remove vs default merge.
references/schemas.md — JSON shape for voice_lint.py, plain-text shape for extract_rules.py. Stable contract for downstream consumers.
scripts/voice_lint.py — validates a BRAND-VOICE.md, walks voice.extends chain, emits chain and merged_stats when inheritance applies. Python 3.7+, no third-party deps.
scripts/extract_rules.py — emits flat testable rules. Resolves voice.extends chain by default. --full (default) includes core_attributes/contexts/source_urls; --legacy emits the v1 minimal output. Consumed by humanize-en -f.
scripts/measure_corpus.py — measures stylometric stats from a prose corpus; --as-sentence-norms emits a sentence_norms dict (or null below the 30-sentence threshold) for extract to use in place of estimated bounds. stdlib only.
scripts/lint_all.py — globs every BRAND-VOICE*.md under a root and lints each. Single-command audit for the parent-change blast-radius problem: a parent edit that breaks N children surfaces as N RED verdicts. CI-friendly; recommended in pre-merge hooks.
Child _replace directives win over parent rules even when --chain shows the parent.scripts/utils.py:resolve_extends_chain walks parent → child; merge applies overrides last. Verify final state with voice_lint.py (chain state is in its JSON output; --chain is a show flag) before merging.
A structurally invalid parent passes chain resolution but fails lint when linted directly. Fix: lint every file in the chain via scripts/lint_all.py, not just the target; block writes if any ancestor is RED.
_replace / _remove only apply to fields in REPLACE_ALLOWED_FIELDS / REMOVE_ALLOWED_FIELDS (scripts/utils.py:334-352). Overrides on other fields (e.g., voice, source_urls, signature_traits) no-op at merge and surface only at lint time. Always run voice_lint.py on the child after override edits.
Chains over 5 hops fail with extends-depth-exceeded (scripts/utils.py:332: MAX_EXTENDS_DEPTH = 5). Fix: keep chains short (≤3 hops); flatten when a child needs more than 2 ancestors.
extract_rules.py path mismatch breaks humanize-en -f silently. Consumers try three candidates and degrade to universal patterns if none resolve. Symlink or copy the script when your install path is non-standard.
