brand-voice
Extracts or builds a falsifiable brand voice profile, backed by verbatim evidence from the brand's real content, that every other content skill reads, so output sounds like the brand instead of a model. Use when starting a project, when copy sounds generic or AI-written, when output must match a specific person or brand, or to check recently published content for voice drift.
boringmarketer · Static checks on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add boringmarketer/boring-marketing-skills --skill brand-voiceAbout this skill
name: brand-voice description: Extracts or builds a falsifiable brand voice profile, backed by verbatim evidence from the brand's real content, that every other content skill reads, so output sounds like the brand instead of a model. Use when starting a project, when copy sounds generic or AI-written, when output must match a specific person or brand, or to check recently published content for voice drift. when_to_use: what's my voice, analyze my brand voice, make this sound like me, voice guide, brand personality, analyze my website, this sounds generic, this sounds like AI, voice drift check, is this on brand, define my voice argument-hint: "[URL to auto-scrape | pasted content to extract | 'drift check' + recent content]" metadata: boring: version: 4.1.1 layer: foundation reads: [positioning, audience] writes: [voice-profile] optional-keys: [FIRECRAWL_API_KEY, EXA_API_KEY]
Brand Voice
Produce a voice profile specific enough that a stranger holding only the profile writes content the brand's audience would attribute to the brand, and tight enough that every line rules something out. The judgment this skill carries: voice claims are evidence-backed and falsifiable, never horoscope; the Voice Test Loop, not interrogation, is how the profile gets accurate. The method is a floor, not a ceiling. When the content in front of you argues for a better path, take it and say why.
Context
Only if ./brand/, ./brands/ or MARKETING.md exists (nothing to load otherwise), run:
python3 ${CLAUDE_SKILL_DIR}/../start-here/scripts/brand_context.py --for brand-voice
(If the script or python3 is unavailable, read the files named in reads: directly. If
none exist, proceed standalone: ask only for what the task strictly requires.)
Positioning shapes stance (a rebel brand and a trusted-advisor brand land differently on every spectrum); audience shapes jargon tolerance and register. Use them as priors, never as substitutes for evidence.
Before writing or updating any profile, load references/voice-extraction-judgment.md.
Its falsifiability rule, exhibit-citation requirement, and competitor-paste test are
the quality bar for every line of the profile.
Modes
| Mode | When | Signal |
|---|---|---|
| Extract | They have content they're proud of | Pasted content, file paths, "analyze these" |
| Build | No crawlable surface exists, or existing content is deliberately being abandoned | "I don't have content" AND no site/socials to mine |
| Auto-Scrape | A crawlable surface exists, whether or not they mentioned it | URL in the message, or a domain/handle findable in brand.yaml / positioning.md / stack.md |
| Drift Check | Saved profile exists; they want recent content audited | "drift check," "is this still on brand," recent posts + existing profile |
Tools pick the mode (extract, don't interview): when the web-scrape capability is present, check for a crawlable surface BEFORE asking anything. The message, then brand.yaml, positioning.md, stack.md for a domain or social handles. Found one → Auto-Scrape is the first pass, even if they said "help me define my voice": crawl, extract, present the drafted profile for correction. Build's questions then fill only the gaps the corpus couldn't answer (usually the anti-model and hard lines). A person with a live website should never face an interview a crawler could have skipped.
If ./brand/voice-profile.md already exists and the request isn't Drift Check, this is
an update: read the existing profile, apply the requested change or re-extract on
new evidence, run the Voice Test Loop on what changed, then save (auto-version per
start-here/references/brand-memory.md; show a 3-line summary of what changed, with no confirmation ceremony).
If no surface is findable and the mode is still ambiguous, ask ONE question: "Do you have published content that sounds like you, or are we building from scratch? (A URL works too: I'll do the research.)"
The six dimensions
Every mode analyzes (or constructs) the same six dimensions. These are the skeleton of the profile.
- Tone: position on five spectrums: Formal↔Casual, Serious↔Playful, Reserved↔Bold, Simple↔Sophisticated, Warm↔Direct. A position is only useful with its mechanism ("casual in syntax, never in punctuation" beats "pretty casual").
- Vocabulary: signature words and coinages, banned words, jargon level and what jargon signals (peer vs seller), profanity policy.
- Rhythm: sentence-length distribution (the mix, not the average), paragraph shape, fragment use, where the prose speeds up and slows down.
- Structure: how pieces open (story / claim / question), how they transition, how they close (CTA / open loop / cold stop), formatting habits.
- Personality: the 3-4 traits that show up as observable behavior in the prose: self-deprecating about what? confident how? teacher or peer? What's the recurring move?
- POV & address: I vs we, how the reader is addressed, the relationship stance (peer slightly ahead / guide / insider / rebel).
Mode: Extract
Input: 3-5 pieces the user considers "most them": posts, emails, transcripts, pages. Under ~1,500 words total, say the corpus is too thin for reliable patterns and offer to Auto-Scrape the gaps or proceed with caveats.
Process:
- Read everything once for gestalt, then once per dimension.
- Pull verbatim exhibits (5-60 words each) as you go. Every pattern you think you see must be pinned to a quote. Target 10-20 exhibits.
- Write the profile (template below), citing exhibit IDs per the falsifiability rule.
- Run the Voice Test Loop.
What separates a good extraction from a mediocre one: the mediocre one describes what the content is ABOUT; the good one describes what the writer DOES. The repeated move, the thing they never do, the habit they probably don't know they have. If you can't name three habits the brand itself would be
Read the complete source ↗