How to Write a CLAUDE.md File for Your Brand (Complete Guide)

What is CLAUDE.md?

CLAUDE.md is a markdown file that lives in your project root and tells AI agents how to behave. Originally designed for Claude Code (Anthropic's CLI tool), the format has become a de facto standard across AI tools. ChatGPT reads it when uploaded. Copilot respects it. Custom agents load it at startup.

Think of it as a system prompt that lives in version control: diffable, reviewable, and shared across your entire team.

Why your brand needs one

Without a CLAUDE.md, every AI interaction starts from zero. The agent has no memory of your brand voice, no knowledge of your color system, no understanding of your messaging hierarchy. Every prompt requires manual context injection. Every output requires manual brand review.

With a CLAUDE.md, agents start from your brand. Your voice is the default, not an afterthought.

The data supports this: companies with documented AI governance frameworks see 65% reduction in content review time and 18% faster product cycles.

The complete CLAUDE.md template

Here is a production-ready template. Copy it, fill in your brand details, and commit it to your repo.

# Brand Guidelines - [Your Company Name]

## Identity
- **Name**: [Company Name]
- **Tagline**: [Your tagline]
- **Industry**: [Your industry/vertical]
- **Audience**: [Primary audience description]
- **Positioning**: [One sentence on what makes you different]

## Voice & Tone

### Default voice
[2-3 sentences describing your default communication style.
Example: Clear, confident, and evidence-led. We use short
sentences. We avoid jargon unless the audience is technical.
We never use superlatives like "best" or "revolutionary."]

### Context-specific tone shifts

#### Support
[How you sound when helping customers.]
- Do: Empathize first, then solve. One apology maximum.
- Don't: Dwell on the problem. Use corporate language.
- Example: "We're sorry about the delay. Here's how to fix it."

#### Marketing
[How you sound when selling.]
- Do: Lead with data. Make claims specific and verifiable.
- Don't: Use hype words. Promise without evidence.
- Example: "87% of teams ship faster in the first month."

#### Technical / Documentation
[How you sound when explaining.]
- Do: Be precise. Use code examples. Assume competence.
- Don't: Be condescending. Over-explain obvious things.
- Example: "Pass the config object to initialize()."

#### Social Media
[How you sound on social platforms.]
- Do: Be conversational. Respond quickly. Show personality.
- Don't: Be stiff. Ignore comments. Use formal language.

## Visual Identity

### Colors
| Role       | Hex       | Usage                           | Avoid                    |
|------------|-----------|----------------------------------|--------------------------|
| Primary    | [#hex]    | CTAs, key links, hero accents   | Body text, backgrounds   |
| Accent     | [#hex]    | Secondary actions, highlights   | Primary CTA buttons      |
| Background | [#hex]    | Page backgrounds                | Text color               |
| Dark       | [#hex]    | Dark sections, footer           | Light-mode text bg       |

### Typography
| Role     | Font               | Size  | Weight   | Usage                |
|----------|-------------------|-------|----------|----------------------|
| Display  | [Font Name]       | 2-4rem| Regular  | H1, hero text        |
| Body     | [Font Name]       | 16px  | 400-500  | Paragraphs, UI text  |
| Code     | [Font Name]       | 14px  | 400      | Code blocks, data    |

### Spacing
- Section padding: [value]
- Component gap: [value]
- Minimum touch target: 48px

## Messaging Hierarchy

### Primary message
[Your core value proposition in one sentence.]

### Supporting messages
1. [Benefit #1 with proof point]
2. [Benefit #2 with proof point]
3. [Benefit #3 with proof point]

### Proof points
- [Statistic or customer result]
- [Statistic or customer result]
- [Award, certification, or social proof]

## Brand Rules

### Always
- [Rule 1: e.g., "Back every claim with data"]
- [Rule 2: e.g., "Use active voice"]
- [Rule 3: e.g., "Include a clear next step"]

### Never
- [Rule 1: e.g., "Never use 'innovative' or 'cutting-edge'"]
- [Rule 2: e.g., "Never make compliance claims without legal review"]
- [Rule 3: e.g., "Never use more than one exclamation mark"]

## Terminology
| Use this          | Not this              |
|-------------------|-----------------------|
| [preferred term]  | [deprecated term]     |
| [preferred term]  | [deprecated term]     |

Section-by-section breakdown

Identity - Who you are

This section grounds the agent. Without it, the agent guesses your industry, audience, and positioning. With it, every output starts from the right context.

Keep it short. Five lines maximum. The agent does not need your company history. It needs your current positioning.

Voice & Tone - How you sound

This is the most important section. Voice is the single biggest source of brand drift in AI-generated content. Without explicit rules, agents default to a generic, slightly corporate tone that sounds like every other company.

The key insight: Voice is not one thing. It shifts by context. Your support voice is different from your marketing voice, which is different from your documentation voice. Document each shift explicitly, with do/don't examples.

Bad voice section:

## Voice
Be friendly and professional.

Good voice section:

## Voice
### Default
Clear, confident, evidence-led. Short sentences. No jargon
unless the audience is technical. Never use "innovative,"
"cutting-edge," or "revolutionary."

### Support
Empathize, then solve. One apology, then action.
- Do: "We're sorry about the delay. Here's the fix."
- Don't: "We sincerely apologize for any inconvenience
  this may have caused to you and your team."

The second version gives the agent specific, actionable constraints. The first gives it nothing.

Visual Identity - How you look

Colors, typography, and spacing are the visual tokens your brand is built on. Document them with roles and usage rules, not just values.

Saying "Primary: #9C4221" is not enough. Saying "Primary: #9C4221 - CTAs and hero accents only. Never for body text on light backgrounds" gives the agent a constraint it can enforce.

For more on design tokens and how agents consume them, see our guide on design tokens for AI.

Messaging Hierarchy - What you say

Your messaging hierarchy is the stack of claims that support your positioning: primary message at the top, supporting messages below, proof points at the base.

Without this, agents invent claims. With this, agents ladder messages correctly: the right claim at the right level of specificity.

Brand Rules - The hard constraints

Rules are the guardrails. They are binary: always do this, never do that. They catch the mistakes that voice guidance alone misses.

Good rules are specific and testable:

• "Never make compliance claims (HIPAA, SOC 2, GDPR) without [legal-reviewed] tag"
• "All statistics must include source and date"
• "Error messages: one sentence maximum, always include a next step"

Terminology - The vocabulary

Every brand has words it prefers and words it avoids. A terminology table ensures consistency across agents, teams, and contexts.

Real-world CLAUDE.md example

Here is a condensed real example for a B2B SaaS company:

# Brand Guidelines - Acme Analytics

## Identity
- **Name**: Acme Analytics
- **Tagline**: Data that decides.
- **Industry**: B2B analytics / business intelligence
- **Audience**: Data teams at mid-market SaaS (50-500 employees)
- **Positioning**: Fastest path from raw data to board-ready insight.

## Voice
### Default
Direct, data-driven, slightly informal. We explain with numbers,
not adjectives. "23% faster" beats "significantly faster" every time.

### Support
Fix first, empathize second. Our users are technical. Respect
their time. Skip pleasantries in bug reports.
- Do: "Fixed in v2.4.1. Update via pip install --upgrade."
- Don't: "Thank you so much for reaching out! We really
  appreciate your patience..."

### Marketing
Numbers lead. Every claim has a source.
- Do: "Teams using Acme ship dashboards 3x faster (2025 survey, n=400)."
- Don't: "Acme is the best analytics platform on the market."

## Visual
| Role      | Hex     | Usage                |
|-----------|---------|----------------------|
| Primary   | #2563EB | CTAs, links          |
| Dark      | #0F172A | Headings, nav        |
| Light     | #F8FAFC | Backgrounds          |

## Rules
### Always
- Cite data sources within the same paragraph
- Use active voice in all customer-facing copy

### Never
- Say "AI-powered" (our users find it meaningless)
- Use exclamation marks in documentation
- Make speed claims without benchmark methodology

How to deploy your CLAUDE.md

For Claude (Claude Code / Claude.ai)

Drop CLAUDE.md in your project root. Claude Code reads it automatically. For Claude.ai, paste the contents into the project instructions.

For ChatGPT

Paste the contents into Settings > Personalization > Custom Instructions. Or upload the file directly to a conversation.

For Cursor

Create a .cursorrules file with the code-specific subset. See CLAUDE.md vs .cursorrules for what goes where.

For GitHub Copilot

Create an AGENTS.md file. Copilot reads it. See our AGENTS.md guide.

For all tools simultaneously

Use BrandMythos. Upload your brand guide once, get CLAUDE.md + .cursorrules + AGENTS.md + system prompts + tokens + knowledge graph. All synced and versioned.

Common mistakes

1. Too vague. "Be professional" means nothing to an agent. Be specific: "Use title case for headings. No contractions in legal copy."

2. Too long. A 5,000-word CLAUDE.md wastes context window. Keep it under 2,000 words. Move reference data to separate files.

3. No examples. Rules without examples are ambiguous. Always include a "Do" and "Don't" for voice guidance.

4. Static. A CLAUDE.md that never changes drifts from reality. Review it monthly. Treat it like code. It deserves pull requests.

5. Missing context shifts. One voice for all contexts is not governance. It is laziness. Document at least 3 context shifts (support, marketing, technical).

Measuring impact

After deploying your CLAUDE.md, track:

• First-pass approval rate: What percentage of AI outputs pass brand review without edits?
• Review cycle time: How long does brand review take per piece of content?
• Brand consistency score: Sample 50 AI outputs/month and score on a 1-5 brand alignment scale.

Teams typically see first-pass approval rates jump from ~40% to ~85% within the first month of deploying a CLAUDE.md.

Start now

The best CLAUDE.md is the one that exists. Start with voice rules and three constraints. Expand as you learn what the agent gets wrong.

Or let BrandMythos do it for you. Try it with your brand. Enter your URL, and we will extract your voice, colors, typography, and rules into a production-ready CLAUDE.md in minutes.