Start free: install Anthropic’s brand-voice plugin (claude plugin install brand-voice@knowledge-work-plugins after adding the anthropics/knowledge-work-plugins marketplace), keep your rules in a project SKILL.md committed to git, and run a humanizer pass on drafts. A voice skill fixes the familiar problem of pasting the style guide in every session: Claude loads the rules whenever the task matches.
This guide covers how to pull a voice out of content you already have, how to store it as a skill, how to write banned and preferred word lists, how to share one voice across a team through git, and how to run voice reviews. It also lists the free and paid skills worth looking at, with install commands checked against each repo’s README on 2026-11-01.
New to skills? Start with what Claude Code skills are and come back.
What a brand voice skill actually is
Short answer: a brand voice skill is a SKILL.md file (plus reference files) that tells Claude how your brand sounds, which words to use and avoid, and how to check a draft. It lives in ~/.claude/skills/ for you alone, or in .claude/skills/ inside a project so everyone who clones the repo gets it.
That is the whole mechanism. There is no model training and no fine-tuning. Claude reads the skill’s short description at the start of a session, and when your request matches (write a landing page, edit this email, review this post), it loads the full instructions and any reference files the skill points to.
This matters for voice work for three reasons:
- Progressive loading. Only the description sits in context until the skill fires. That means a long voice guide with 40 examples costs you nothing on unrelated tasks.
- Reference files. You can split a voice into
SKILL.md(the rules) and files likereferences/examples.mdorreferences/word-list.md. Claude reads them when the instructions say to. - It is plain text. You can diff it, review it in a pull request and roll it back. Try doing that with a chat setting.
Here is how a voice skill compares with the other places people put voice instructions:
| Where the voice lives | Persists across sessions | Shared with a team | Version history | Loads only when relevant |
|---|---|---|---|---|
| Pasted into each prompt | No | No | No | n/a |
| Claude.ai Project instructions | Yes | Inside that Project | No | No, always on |
CLAUDE.md in a repo | Yes | Yes, via git | Yes | No, always on |
Personal skill in ~/.claude/skills/ | Yes | No | Only if you track it | Yes |
Project skill in .claude/skills/ | Yes | Yes, via git | Yes | Yes |
| Plugin from a marketplace | Yes | Yes, if everyone installs it | Through the plugin’s repo | Yes |
The takeaway: if more than one person writes for the brand, the project skill or a plugin is the right home. If it is just you, a personal skill is fine.
How to extract a brand voice from content you already have
Short answer: collect 10 to 30 pieces of your best existing writing, ask Claude to describe the patterns it finds (sentence length, vocabulary, point of view, what you never say), then turn that description into rules and examples. You edit the result. Claude drafts the guide; a human approves it.
Most teams skip this step and write aspirational adjectives instead (“friendly, bold, trustworthy”). Those words do almost nothing for a model, because every brand claims them. Patterns pulled from real copy do much more.
Step 1: pick the right source material
Not all of your content is on-voice. Pick pieces that a senior person would point at and say “that is us.”
- Homepage and top landing pages
- Your three to five best performing emails (you know which ones; no stat needed)
- Founder posts or newsletters people reply to
- Support macros that customers praise
- Sales call transcripts, if your brand voice is conversational
Skip anything written by an agency that never quite got it, and anything already AI-drafted. You do not want the model learning from its own habits.
Step 2: run a structured analysis
Put the files in a folder (for example voice-samples/) and ask Claude Code to analyze them against a fixed checklist. A fixed checklist keeps the output comparable when you rerun it next quarter.
| Dimension | What to ask Claude to report |
|---|---|
| Sentence shape | Average length, how often fragments appear, use of questions |
| Point of view | ”We” vs “you” vs third person, and where each appears |
| Vocabulary | Recurring words, product terms, words that never appear |
| Formality | Contractions, slang, jargon level |
| Rhythm | Paragraph length, use of lists, how sections open |
| Claims | How the brand states proof, hedges, or avoids hype |
| Humor | Present or absent, and what kind |
Ask for quoted evidence for every pattern. “Short sentences” means nothing. “Most homepage sentences run under 12 words, for example: ‘Ship the page. Fix it later.’” is a rule you can test.
Step 3: write “we are / we are not” pairs
Anthropic’s brand-voice plugin (built by Tribe AI, covered below) uses a “We Are / We Are Not” framework, and it is a useful format whatever tool you use. Each pair draws a boundary:
- We are direct. We are not curt.
- We are plain-spoken. We are not dumbed down.
- We are confident about what we have shipped. We are not boastful about what we have not.
Step 4: keep voice and tone separate
Voice stays constant. Tone shifts by channel. A cold email and a help doc should sound like the same company in different moods. The thatrebeccarae/claude-marketing brand-voice-guidelines skill and Anthropic’s plugin both make this split, and your skill should too. A small tone table (channel, formality, energy, length) handles most of it.
If you want a tool to do the discovery for you, two options do it out of the box: Anthropic’s brand-voice plugin (/brand-voice:discover-brand) searches connected sources like Notion, Google Drive and Slack, and the brand-voice-learner skill listed on skills.sh analyzes content you supply. I have not verified the quality of either one’s output on real brands, so treat them as a first draft.
Voice guides as skill reference files: a sample SKILL.md
Short answer: keep SKILL.md short (rules, workflow, when to load references) and push the bulky material (examples, word lists, tone tables) into a references/ folder. Claude reads references only when the instructions tell it to, so the guide can be long without slowing anything down.
A folder layout that works:
.claude/skills/brand-voice/
├── SKILL.md
└── references/
├── voice-rules.md # we are / we are not, tone by channel
├── word-list.md # banned, preferred, product terms
├── examples.md # 5-10 before/after rewrites
└── claims-policy.md # what we can and cannot claimHere is a sample SKILL.md you can copy and fill in. Replace the bracketed parts.
---
name: brand-voice
description: Write and edit any customer-facing copy in [Brand]'s voice. Use when drafting or rewriting landing pages, emails, ads, social posts, docs or sales copy for [Brand], or when the user says "on-brand", "in our voice" or "voice check".
---
# [Brand] voice
You are writing as [Brand]. Read references/voice-rules.md and
references/word-list.md before writing anything. For rewrites, also
read references/examples.md.
## The voice in one paragraph
[Brand] talks like a senior operator explaining something to a smart
peer: short sentences, concrete nouns, no hype. We say what the product
does, what it does not do, and what to do next.
## Rules
1. Lead with the answer or the outcome in the first sentence.
2. Second person ("you") for instructions. "We" only for things the
company actually did.
3. Average sentence under 18 words. Vary length.
4. Contractions are fine. Exclamation marks: none in body copy.
5. Numbers as digits. Every number needs a source in
references/claims-policy.md, or it gets cut.
6. Never use a word from the banned list. Prefer the preferred term.
7. No em dashes. Use a comma, a full stop or brackets.
## Tone by channel
| Channel | Formality | Energy | Typical length |
|-----------|-----------|--------|----------------|
| Homepage | Medium | High | Short |
| Email | Low | Medium | 80-150 words |
| Docs | Medium | Low | As needed |
| LinkedIn | Low | Medium | 100-200 words |
## Workflow
1. Draft in the voice.
2. Run the self-check below.
3. Output the draft, then a short "voice notes" list of any rule you
bent and why.
## Self-check before output
- [ ] No banned words (check references/word-list.md)
- [ ] First sentence carries the point
- [ ] No unsourced claims
- [ ] Reads like the samples in references/examples.mdTwo details make this work. First, the description field lists real trigger phrases, because that is what Claude matches against when deciding to load the skill. Second, the self-check is part of the skill, so every draft ships with its own review. For more on structure, see how to build a Claude Code skill.
Banned and preferred words: the highest-return part of any voice skill
Short answer: a word list is the most testable part of a voice guide. Keep three lists: banned words (never use), preferred terms (use this, not that) and product terms (exact spelling and capitalization). Put them in a reference file and have the skill check every draft against them.
Adjectives describing a voice are hard to enforce. A word list is binary. Either “leverage” is in the draft or it is not, which also means a script can check it.
What goes on each list
| List | Purpose | Example entries |
|---|---|---|
| Banned | Words that sound generic, hypey or off-brand | next-gen, supercharge, best-of-breed, ninja, rockstar |
| Preferred | Swap a weak or off-brand term for yours | ”customers” not “users”; “set up” not “onboard”; “plan” not “tier” |
| Product terms | Exact names and spelling | ”Claude Code” (two words, both capitalized), your feature names |
| Claim words | Words that need proof attached | ”fastest”, “guaranteed”, “proven”, “#1” |
A few rules from teams that maintain these lists:
- Give the reason. “Banned: ‘revolutionary’ (every competitor says it)” helps Claude generalize to similar words you did not list.
- Give the replacement. A ban without an alternative produces awkward workarounds.
- Keep it under 50 banned entries. Past that, nobody maintains it and Claude starts over-correcting.
- Review it quarterly. Language drifts. New AI tells appear.
Pair it with a humanizer
Generic AI phrasing is a voice problem too. Humanizer skills such as blader/humanizer (63,282 GitHub stars, last push 2026-10-28) flag patterns like inflated phrasing, rule-of-three lists and em dash overuse, based on Wikipedia’s “Signs of AI writing” guide. Run your voice skill to draft and a humanizer to clean up. More on that in Claude Code skills for humanizing AI writing.
Enforcing one voice across a team: project skills in git
Short answer: commit the voice skill to .claude/skills/brand-voice/ in a shared repo. Everyone who clones it and runs Claude Code in that folder gets the same skill. Changes go through pull requests, so the voice has an owner, a history and a review step.
This is where Claude Code beats chat tools for brand work. A chat setting is per person. A file in git is per team.
A simple setup
- Create a repo (for example
marketing-content) or use the one your site lives in. - Add
.claude/skills/brand-voice/with theSKILL.mdand references above. - Add a
CODEOWNERSentry so the brand lead approves changes to that folder. - Ask writers to open Claude Code from the repo root. The skill is available automatically.
- Tag releases when the voice changes (“voice v2: dropped exclamation marks”) so you can tell which drafts used which rules.
When to use a plugin instead
If your team works across many repos, a plugin is easier than copying the folder around. You publish the skill in a repo with a marketplace file, and teammates run /plugin marketplace add your-org/brand-voice and /plugin install. The trade-off is that people must install and update it themselves, while a project skill arrives with git pull.
| Approach | Best for | Weak spot |
|---|---|---|
Project skill in .claude/skills/ | One content repo, small team | Copying to other repos |
| Private plugin marketplace | Many repos, larger team | Each person installs and updates |
| Personal skill | Solo writers, freelancers | Nothing shared |
One honest limit: a skill shapes what Claude writes. It cannot stop someone from editing a draft off-voice afterward, and it does not log usage. Markup AI makes a version of this argument in its post on why a Claude skill is not a full brand control system. That is fair. If you need audit trails across hundreds of writers, you need something beyond a skill. For most teams, a reviewed skill in git plus a review step is plenty.
See how to install Claude Code skills for the install basics.
Voice review and audit skills
Short answer: a voice review skill takes an existing draft and scores it against your guide, flagging each deviation with a severity and a suggested rewrite. Use it as the last step before anything ships, and run it in batches to audit older content.
Drafting and reviewing are different jobs. A drafting skill nudges the writing. A review skill reads finished copy cold and reports what is wrong. Keeping them separate gives you a second pass that is not grading its own homework.
What a good review skill outputs
- A pass/fail or score per rule in your guide
- Each issue quoted, with severity (blocker, should fix, nit)
- A specific before/after rewrite per issue
- Claims that need a source, flagged separately
- Banned words found, with line references
Anthropic’s Marketing plugin includes a brand-review skill that does close to this: per its SKILL.md, it reviews content “against your brand voice, style guide, and messaging pillars, flagging deviations by severity with specific before/after fixes,” and also screens for unsubstantiated claims and missing disclaimers.
Running a content audit
To audit a back catalog, point Claude Code at a folder of posts and ask it to run the review skill on each one, writing results to a single table. Sort by severity and fix the blockers first. Pair this with Claude Code skills for editing if the fixes go beyond voice.
Anthropic’s brand-voice plugin, verified
Short answer: yes, it exists. It lives in anthropics/knowledge-work-plugins under partner-built/brand-voice, was built by Tribe AI as a Cowork launch partner, and ships three commands: /brand-voice:discover-brand, /brand-voice:generate-guidelines and /brand-voice:enforce-voice. It is free.
I checked the repo on 2026-11-01 (34,325 stars, last push 2026-11-01). What the README says it does:
- Discovery. Searches connected platforms (Notion, Confluence via Atlassian, Box, Microsoft 365, Figma, Gong, Granola, plus Google Drive and Slack as native integrations) for brand material.
- Guideline generation. Produces voice pillars, tone parameters, a “We Are / We Are Not” framework and terminology standards.
- Enforcement. Writes new content against those guidelines, with tone flexing by context.
- Open questions. When sources conflict, it lists questions with a recommendation for you to confirm or override.
Two caveats from the README itself. It is written for Claude Cowork first, and it says “Brand Voice currently works at the individual level, team-wide enforcement is coming soon.” For Claude Code, you can copy its settings file to .claude/brand-voice.local.md in your project, per the README.
Install (from the repo README):
claude plugin marketplace add anthropics/knowledge-work-plugins
claude plugin install brand-voice@knowledge-work-pluginsThe marketplace file lists the plugin as brand-voice. The separate marketing plugin in the same marketplace (claude plugin install marketing@knowledge-work-plugins) contains the brand-review skill.
Do not confuse it with Anthropic’s brand-guidelines skill in anthropics/skills, which is mainly about Anthropic’s own visual identity (colors and type), not your voice.
The best brand voice skills: free and paid
Short answer: most teams should start free. Anthropic’s brand-voice plugin handles discovery and enforcement, coreyhaines31/marketingskills gives you a shared product context plus copy-editing, and a humanizer cleans up AI tells. Paid packs make sense only if you want a broad marketing library in one purchase.
Star counts and last-push dates come from the GitHub API on 2026-11-01. Install commands were checked against each README.
Free skills
1. Anthropic brand-voice plugin (Tribe AI)
- Source: github.com/anthropics/knowledge-work-plugins (partner-built/brand-voice)
- Price: free
- Does: discovers voice from connected tools, writes guidelines, enforces them in drafts
- Install:
claude plugin marketplace add anthropics/knowledge-work-pluginsthenclaude plugin install brand-voice@knowledge-work-plugins
2. Anthropic Marketing plugin (brand-review skill)
- Source: same repo,
marketing/ - Price: free
- Does: reviews drafts against your voice and style guide with severity and fixes
- Install:
claude plugin install marketing@knowledge-work-plugins
3. coreyhaines31/marketingskills (57,803 stars, last push 2026-10-31)
- Price: free
- Does: the
product-marketingskill builds a shared context doc (product, audience, positioning) that every other skill reads first. That doc is a good home for voice notes. Also includescopywritingandcopy-editing. - Install:
npx skills add coreyhaines31/marketingskillsor/plugin marketplace add coreyhaines31/marketingskillsthen/plugin install marketing-skills
4. alirezarezvani/claude-skills (29,110 stars, last push 2026-10-31)
- Price: free
- Does: a
brand-guidelinesskill in the marketing set, with tone-of-voice guidance - Install:
/plugin marketplace add alirezarezvani/claude-skillsthen/plugin install marketing-skills@claude-code-skills
5. thatrebeccarae/claude-marketing (188 stars, last push 2026-05-14)
- Price: free
- Does:
brand-voice-guidelinesskill that separates voice from tone across channels - Install:
git clone https://github.com/thatrebeccarae/claude-marketing.gitand copy the skill folder into~/.claude/skills/
6. luongnv89/claude-howto (46,346 stars, last push 2026-10-29)
- Price: free
- Does: a tutorial repo that includes a worked
brand-voiceexample skill. Useful as a template more than a finished tool. - Install: clone the repo and copy the skill folder
7. rampstackco/claude-skills (1,013 stars, last push 2026-10-27)
- Price: free
- Does: includes a
brand-voiceskill (listed on skills.sh) - Install:
/plugin marketplace add rampstackco/claude-skillsthen/plugin install rampstack-skills@rampstack
8. blader/humanizer (63,282 stars, last push 2026-10-28)
- Price: free
- Does: strips AI writing patterns; pairs with any voice skill
- Install:
/plugin marketplace add blader/humanizerthen/plugin install humanizer@humanizer
9. ComposioHQ/awesome-claude-skills (78,504 stars, last push 2026-09-18)
- Price: free
- Does: a curated list, not a skill. Good for finding more voice and writing skills.
Directories such as claudemarketplaces.com, skills.sh, claudskills.com and skillselion.com also list brand voice skills (for example brand-voice-coach and brand-voice-learner). Listings change quickly, and I have not reviewed every one, so check the source repo before installing.
Paid packs
None of these is a dedicated brand voice product. They are broad marketing or SEO libraries that include copy and editing skills. Prices were checked on the live pages on 2026-11-01.
| Product | Where | Price | What you get | Voice-relevant? |
|---|---|---|---|---|
| Master Marketing Skills Bundle (Inflectual) | Gumroad: inflectual.gumroad.com/l/master-marketing-claude-skills-bundle | $167 one-time | 167 marketing skills, per the listing | Broad copy coverage; no dedicated voice skill confirmed |
| Ultimate Claude SEO & AEO Skills Pack | Contra: contra.com/products/fE9zVnna-the-ultimate-claude-seo-and-aeo-skills-pack | $39 (shown against $79) | SEO and AEO writing and audit skills | Indirect; SEO-focused |
| SuperSkills | superskills-one.vercel.app | $50 one-time | Curated bundle of Claude Code skills across categories | General; check the category list first |
Left out on purpose: Claude Skills 360 lists $39 one-time, but its pricing section runs a countdown timer computed from the visitor’s clock rather than a real deadline. That is the fake-urgency pattern this site does not recommend, so it is not in the table.
Free vs paid at a glance
| Free (GitHub, Anthropic) | Paid packs | |
|---|---|---|
| Dedicated brand voice skill | Yes (Anthropic plugin, several community skills) | Not confirmed in any pack listed |
| Source visible before install | Yes | Usually after purchase |
| Team sharing via git | Yes | Depends on license; check terms |
| Updates | Through the repo | Vendor-dependent |
| Best for | Teams that want to own their voice files | Buyers who want a wide library at once |
My recommendation: a voice skill is mostly your own content. No pack can ship your word list. Start with a free skill, write your references, and buy a pack only for the other marketing work you need. For broader picks, see Claude Code skills for marketers and Claude Code skills for copywriters.
Common mistakes with brand voice skills
Short answer: the usual failures are vague adjectives instead of rules, no examples, a description that never triggers, and one giant file nobody maintains.
- Adjectives only. “Warm and confident” gives Claude nothing to check. Add rules and examples.
- No before/after examples. Five good rewrites teach more than five pages of rules.
- Weak description field. If the description does not mention the tasks writers actually request, the skill never loads.
- Mixing voice and claims. Keep a separate claims policy so the voice skill does not “sound confident” by inventing numbers.
- No owner. Without a CODEOWNERS entry, the word list rots.
- Training on AI drafts. Extracting voice from AI-written copy bakes in AI habits.
How to start this week
- Gather 10 to 30 on-voice samples into
voice-samples/. - Ask Claude Code to analyze them against the dimension table above, with quotes.
- Edit the output into
voice-rules.md,word-list.mdandexamples.md. - Copy the sample
SKILL.md, fill in your brand, and commit it to.claude/skills/brand-voice/. - Install a review skill (Anthropic’s
brand-reviewor your own) and a humanizer. - Run the review on your last 10 published pieces. Fix blockers, then update the word list with what you learned.
Writing for specific channels next? See skills for landing page copy, skills for email marketing and skills for LinkedIn posts.
FAQ
Can Claude learn my brand voice?
Not by training. It follows instructions and examples you give it. A skill makes those instructions load automatically each time, which gets you most of the consistency people mean by “learning.”
How do I train Claude to write in my voice?
Collect your best samples, have Claude describe the patterns with quotes, turn that into rules and a word list, and save it as a skill with before/after examples.
What is the difference between a brand voice skill and CLAUDE.md?
CLAUDE.md loads in every session in that project. A skill loads only when the task matches its description. Put short global rules in CLAUDE.md and the full voice guide in a skill.
Does Anthropic have an official brand voice plugin?
Yes. The brand-voice plugin in anthropics/knowledge-work-plugins was built by Tribe AI. It is free and adds discover, generate and enforce commands.
Does the Anthropic brand-voice plugin work in Claude Code or only Cowork?
It is written for Cowork first. It installs through the same plugin marketplace Claude Code uses, and the README describes a per-project settings file for .claude/. Connector-based discovery needs those MCP connections set up.
Is Anthropic’s brand-guidelines skill the same thing?
No. The brand-guidelines skill in anthropics/skills is mainly about Anthropic’s own visual identity. For your voice, use the brand-voice plugin or your own skill.
Where do I put a brand voice skill?
~/.claude/skills/brand-voice/ for personal use, or .claude/skills/brand-voice/ in a repo to share it with your team.
How do I share one voice across my whole team?
Commit the skill to a shared repo, protect the folder with CODEOWNERS, and have writers run Claude Code from that repo. For many repos, package it as a plugin.
How many writing samples do I need?
There is no published minimum. Ten to 30 strong pieces across your main channels is a practical range; quality matters more than count. Treat this range as a working guideline, unconfirmed by any study.
Should I include banned words in my voice skill?
Yes. A banned list with replacements is the easiest rule to enforce and to check with a script.
Can a skill check a draft against my brand voice?
Yes. A review skill such as the brand-review skill in Anthropic’s Marketing plugin scores a draft against your guide and suggests fixes by severity.
How do I stop Claude from sounding like AI?
Combine your voice skill with a humanizer skill and a banned list of common AI phrases. Then read the result yourself.
Are paid Claude skill packs worth it for brand voice?
Not on their own. None of the paid packs checked here confirms a dedicated voice skill, and your voice files have to come from your content anyway.
Can I use multiple brands in one setup?
Yes. Create one skill per brand (for example voice-acme and voice-beta) with descriptions that name the brand, or keep each brand’s skill in its own repo.
How often should I update a voice skill?
Review it when the brand changes, and check the word list roughly every quarter. Version it in git so you know which drafts used which rules.
Sources and methodology
Checked on 2026-11-01: GitHub API (stars, last push) and READMEs for anthropics/knowledge-work-plugins, coreyhaines31/marketingskills, alirezarezvani/claude-skills, thatrebeccarae/claude-marketing, luongnv89/claude-howto, rampstackco/claude-skills, blader/humanizer and ComposioHQ/awesome-claude-skills. Paid prices from the live pages and Gumroad price metadata. Markup AI’s argument from markup.ai/blog/claude-skill-and-brand-voice/. Listings on claudemarketplaces.com, skills.sh and claudskills.com. No paid pack was purchased or tested.
