All posts

Claude Skills for brand voice: build it once, use it everywhere.

If you only ever build one Claude Skill, build this one. How to extract the voice properly, what goes in the SKILL.md, and how it loads itself into every conversation you have.

Of all the Skills a marketer can build, the Brand Voice Skill returns the most. It is the one you touch every single day, because almost everything you ask Claude to write is external-facing copy that has to sound like the brand. Build it once and you stop paying the tax: pasting the voice guide into a fresh conversation, paraphrasing it slightly wrong, watching the output drift a little further from on-brand each time. The voice stops depending on your memory and starts living in the tool.

This is the discipline-specific companion to the complete Claude Skills guide. That post covers what a Skill is and when it earns its place. This one is about one Skill done right: the one that pays back the most.

Step one: extract the voice properly (this is the real work)

The Skill is plumbing. The voice guide it points at is the actual deliverable, and most voice guides are useless to a model because they are written for a brand workshop, not for Claude. A guide that says "we are confident, approachable, and human" gives Claude nothing to act on. Every brand says that. None of it changes a single sentence of output.

What works is example-based, not adjective-based. Specifically, two moves that an adjective list cannot do.

Paired contrast. Never name an attribute alone. Name it against the thing it is constantly mistaken for. "Confident, not arrogant. Specific, not technical. Plain, not flat. Warm, not chummy." The contrast is where the instruction actually lives. "Confident" is a vibe. "Confident, not arrogant" is a boundary Claude can hold a draft against.

Do/don't pairs from real samples. Take a real on-brand sentence and the off-brand version of the same sentence side by side. Six of these teach the model more than three pages of description, because they show the line instead of describing it. This is exactly what the free Voice and Tone Extractor produces: feed it your best existing copy and it returns paired contrasts and do/don't pairs, formatted as a file. The full extraction walkthrough covers the prompt in depth, and you can see a real voice guide it produced before you run it yourself.

Step two: what goes in the SKILL.md

The Skill file does two jobs. The YAML frontmatter at the top tells Claude when to load the Skill. The body tells it what to do once loaded. For a marketing Skill, three sections handle the body: what it knows, what it does, what it refuses.

The knows section points at the voice guide and names the structural moves. The does section is the application logic: apply the rhythm, cite a do/don't pair when a draft hits a don't, suggest the revision. The refuses section is the part most people skip, and it is the part that protects the voice.

Step three: the worked SKILL.md

Here is the full file. Save it as SKILL.md in a folder named brand-voice, drop your extracted guide in the same folder as voice-guide.md, and install it. The 20-minute build walkthrough has the install steps for both Claude Desktop and Claude Code.

---
name: brand-voice
description: Use this Skill whenever the user is writing,
  drafting, rewriting, or critiquing external-facing content
  (emails, social posts, ads, web copy, blog drafts, scripts,
  landing pages) where it needs to sound like the brand.
  Triggers on "write," "draft," "rewrite," "in our voice,"
  "on-brand," "make this sound like us," or any request to
  produce or revise copy for the company. Loads voice-guide.md
  and applies the paired contrasts and do/don't pairs to every
  output. Does not apply to internal notes or raw analysis.
---

# Brand Voice

## What this Skill knows

The full brand voice guide is in voice-guide.md (this folder).
The voice is defined by two example-based structures, not
adjectives:

1. Paired contrast: every attribute named against its
   near-miss (e.g. "confident, not arrogant. specific, not
   technical. plain, not flat.")
2. Do/don't pairs: real on-brand sentences beside the
   off-brand version of the same sentence. See voice-guide.md
   for the 6 pairs that define on-brand vs off-brand.

## What this Skill does

When loaded, apply the voice to every piece of output:
- Hold every sentence against the paired contrasts
- When a draft hits the "don't" side of a pair, name the
  pair and rewrite toward the "do" side
- If no register is specified, use the brand's standard
  register from voice-guide.md
- Match the rhythm in the samples: short sentences,
  concrete nouns, no hedging

## What this Skill refuses

- Do not invent voice attributes that are not in
  voice-guide.md
- Do not "elevate," "punch up," or "make it pop." The voice
  is the voice. Stay inside it.
- Do not use these off-brand words: delve, leverage,
  unlock, transform, best-in-class, resonate, robust,
  seamless, game-changer, in today's world
- If the user asks for output that contradicts the voice
  ("make it more corporate," "sound more excited"), ask
  before doing it. Do not quietly override the guide.

The refuse section is the whole game

Most people write the knows and does sections, ship the Skill, and wonder why the voice still drifts. The drift comes from the two things a positive instruction cannot stop, which is why the refuse section carries the weight.

Refuse to "elevate" the voice. This is the single most common failure. Asked to write in your voice, Claude often produces something slightly grander than your voice: a bit more polished, a bit more agency-deck. It is trying to help. The result is a brand that sounds like every other brand that asked a model to make it sound good. The instruction "do not elevate, the voice is the voice, stay inside it" is what holds the line. Your voice is a target to hit, not a starting point to improve on.

Ban specific off-brand words. A blanket "avoid jargon" does nothing. A named list does everything. Put the actual words in the file: the ones your brand never uses, the ones that signal someone outside the brand wrote the copy. When the banned word is literally in the Skill, Claude will not reach for it, and you will not have to catch it in the edit. This is the difference between a voice guide that describes a preference and a Skill that enforces a rule.

How it loads everywhere, on its own

Here is the payoff, and the reason this beats a Project or a saved prompt. Once installed, the Brand Voice Skill is not pinned to one workspace. It travels. Open a fresh conversation with no Project attached and ask for a launch email: the Skill loads, the voice applies. Open your Q3 campaign Project and draft ten ad variants: the Skill loads there too, on top of the Project's campaign context. Brief an agency from your phone: same voice, no setup.

That is the structural advantage over keeping the voice in a Project. A Project's context only applies inside that Project. A Skill applies across all of them, because it describes how you write, not what you are working on this week. Voice is always true, so it belongs in the layer that is always present. The pillar guide has the full Skill vs Project vs prompt breakdown if you want the decision rule.

The honest caveat

This Skill is exactly as good as voice-guide.md. Wrap a vague, adjective-soup guide in perfect YAML and you get a Skill that reliably produces vaguely on-brand mush. The frontmatter does not fix a weak guide. So spend your time where it pays: extract the voice properly first, with real samples and real contrasts, then wrap it. Ten minutes on the Skill, an hour on the guide. The ratio tells you where the work is.