Brand Guidelines for AI Tools: Ship Your Brand as Markdown, Not a PDF

To make AI tools follow your brand, give them your brand guidelines as a small set of plain-text Markdown files with explicit rules: exact color values and the pairs that pass contrast checks, a type scale, a spacing scale, component rules, and the words you use and avoid. Attach the files wherever each tool keeps standing instructions, and test the output by running the same prompt with and without them. A 60-page PDF brand book is written for people to browse; Markdown files are written to be followed.

Teams now make decks, one-pagers, landing pages and social posts with Claude, ChatGPT, Figma Make, Canva AI, v0 and Lovable. This guide comes from a real project. An enterprise AI client's team was designing decks with an AI design tool, so instead of pointing them to a brand book, we sent the brand primitives (colors, type, spacing, components and tone) as Markdown files the tool could follow. Below is the method, with a template for each file, written for a fictional company. Tool details are as of October 2026.

Why don't PDF brand books work with AI tools?

They are written for people to look at, not for a model to follow. General assistants can read a PDF, but brand books keep their values in images, run long, and describe the brand in feel words a model can't check.

• Values live in pictures. Swatches, type specimens and layout grids are usually artwork. The model may read the caption and miss the hex code, or never see the grid.
• Too much text around too few rules. Most pages are story and inspiration. Every extra page competes with the dozen rules that matter. Figma's guidance for Figma Make makes the same point: many short, focused files keep the context from filling up with irrelevant content.
• Feel words instead of rules. "Confident but approachable" and "generous white space" mean something to a designer. To a model they mean the average of everything it has seen. "96px between sections" is a rule it can follow and you can check.

Figma-only guidelines have a different gap. A frame shows the result, not the rule. A tool that imports one approved slide sees where the title sits, not why, and not which colors must never touch.

What should AI-ready brand guidelines include?

An index file plus six topic files, each short enough to paste into a prompt, written as rules with exact values and do and don't examples. Use the same token names in every file, so "larkspur" means the same blue in colors.md, components.md and your code.

The examples below are for Larkspur, a fictional company that sells payment reconciliation software to finance teams.

index.md

The index tells the tool how to use the set. Put the rules that apply everywhere here. Figma Make, for example, reads its top-level guidelines file first and opens the others when it needs them.

## Larkspur brand guidelines for AI tools
Version 1.3, 2026-10-14. Owner: brand lead, brand@larkspur.example

Read in this order: brand.md, colors.md, typography.md,
layout.md, components.md, copy.md.

- Use only the tokens, fonts and components defined here.
- If a rule is missing, pick the simplest option and say what you assumed.
- If two files conflict, the more specific one wins.
- Never invent logos, customers, metrics or quotes.

## Changelog
- 1.3, 2026-10-14: one signal element per slide, max.
- 1.2, 2026-09-02: chart rules added to components.md.

brand.md

Who you are in a paragraph, then tone of voice as examples. Pairs of do and don't lines show tone better than a list of adjectives.

## Brand
Larkspur matches payments to invoices automatically for finance
teams at mid-market companies. Buyers: CFOs and controllers.
Users: accounts payable teams.

Positioning: close the books faster, with fewer manual checks.

Tone: calm, exact, plain.
- Do: "Match 12,000 payments overnight."
- Don't: "Revolutionize your finance stack!"
- Do: name the problem, then the result.
- Don't: use exclamation marks or jokes about accountants.

colors.md

Every color as a token with a hex value and a job. Write the contrast ratios into the file, so the tool doesn't guess and a reviewer can check. WCAG 2.2 asks for at least 4.5:1 for normal text and 3:1 for large text.

## Colors
| Token    | Hex     | Role                                   |
|----------|---------|----------------------------------------|
| ink      | #121A2B | Text, dark backgrounds                 |
| paper    | #FFFFFF | Default background                     |
| mist     | #F3F5F8 | Section and card backgrounds           |
| slate    | #5A6372 | Secondary text                         |
| larkspur | #3346C2 | Primary buttons, links, chart series 1 |
| signal   | #F5B83D | One highlight per slide or screen      |

Text pairs that pass 4.5:1
- ink on paper 17.4, ink on mist 15.9, ink on signal 9.8
- paper on larkspur 7.5, slate on paper 6.1, slate on mist 5.6

Never
- signal text on paper (1.8) or mist (1.6)
- larkspur text on ink (2.3)
- signal and larkspur as text and background (4.2)

If your site or app uses CSS custom properties, add the same tokens in that form, so tools that write code get the exact names your codebase uses.

:root {
  --color-ink: #121a2b;
  --color-paper: #ffffff;
  --color-mist: #f3f5f8;
  --color-slate: #5a6372;
  --color-larkspur: #3346c2;
  --color-signal: #f5b83d;
}

typography.md

Families, weights and a size scale for each format the team makes. Name fallbacks: if a tool doesn't have your font, it should know what to use instead of picking its own.

## Typography
Display: Source Serif 4, 600. Fallback: Georgia.
Text and UI: Inter, 400 and 600. Fallback: Arial.
No other families or weights.

| Role    | Web  | Slide (1920×1080) | Line height |
|---------|------|-------------------|-------------|
| Display | 56px | 88px              | 1.05        |
| H2      | 36px | 56px              | 1.15        |
| H3      | 24px | 36px              | 1.25        |
| Body    | 18px | 28px              | 1.5         |
| Caption | 14px | 20px              | 1.4         |

- Sentence case for every heading. No all-caps text.
- Slide titles: 12 words max. Web lines: 70 characters max.

layout.md

Numbers, not adjectives: a spacing scale the tool must pick from, the grid for each format, corner radius and imagery rules.

## Layout
Spacing scale (px): 4, 8, 12, 16, 24, 32, 48, 64, 96, 128. No other values.

Web: 12 columns, 24px gutter, 1200px max width, 96px between sections.
Slides: 1920×1080, 96px outer margin, 12 columns, 24px gutter.
Radius: 4px inputs, 8px buttons and cards. No pill shapes.

Imagery
- Real product screens, or abstract shapes in brand colors.
- No stock photos of people at laptops. No 3D renders.
- No text over busy images.

components.md

The parts the team reuses, each with when to use it and its limits. For decks, describe each slide template as a component.

## Components
Primary button: larkspur fill, paper text, 8px radius, 12px 24px padding.
One primary button per screen or slide.
Secondary button: paper fill, ink text, 1px slate border.
Card: mist background, 8px radius, 32px padding, no shadow.

Charts
- Series colors in order: larkspur, ink, slate. Signal for one key value.
- Label values on the chart. No legend under 4 series.
- No 3D charts. No pie charts with more than 4 slices.

Slide templates
- Title: display headline, one-line subtitle, logo bottom left.
- Statement: one sentence, 12 words max, centered.
- Data: a title that states the takeaway, one chart, a source line.
- Never more than 3 bullets on a slide.

copy.md

Words and formatting. This is the file that keeps AI-drafted copy sounding like you; we wrote more about that in should you use AI to write your website copy?

## Copy
Use: match, reconcile, close, exception, finance team.
Avoid: revolutionize, AI-powered, best-in-class, synergy, solution.

- Sentence case headings. No exclamation marks.
- Numerals for every number. Percent with no space: 40%.
- Dates: 14 Oct 2026. Money: $1.2M.
- Always "Larkspur", never "LarkSpur".
- Product names: Larkspur Match, Larkspur Close.

How do you use brand Markdown files in Claude, ChatGPT, Gemini, Figma Make and v0?

Put them where each tool keeps standing context: a project, a custom assistant, a guidelines folder or a saved instruction. Where there is no such place, paste the files a task needs at the top of the prompt. As of October 2026:

• Claude. Upload the files to a project's knowledge and add one line to the project instructions: "Follow index.md for every design and writing task." Anthropic says Projects are available on all plans, including free.
• ChatGPT. Projects hold files and instructions, and project instructions override your global custom instructions. For a custom GPT, OpenAI's guidance is that knowledge is for reference material and rules belong in the instructions, so paste index.md into the instructions and attach the other files as knowledge.
• Gemini. Add the files under Knowledge when you create a Gem. Google says Gems move to skills on personal accounts from November 2026 and on Workspace accounts in 2027, with supported files carried over.
• Figma Make. Each Make file has a guidelines folder, and Make reads Guidelines.md first. Figma recommends short imperative rules, many small files, and named tokens instead of hard-coded values. Your index.md becomes Guidelines.md; the other files sit beside it.
• v0. Instructions are reusable prompts saved to your account, which suits the core rules from index.md. v0's design systems take GitHub repos, Figma frames, docs links and attachments as sources.
• Canva AI, Lovable and other tools. Check whether the tool has a place for standing instructions or knowledge. If not, paste brand.md and copy.md for a post, or colors.md and components.md for a chart.

For teams that build in code, keep the files in a brand folder in the website or app repo, next to the tokens and components. Changes go through the same review as code, and coding agents find them like any other file. It is the same single-source idea we use for design systems that scale.

Keep each file short; if one grows past two pages, split it. Put a version and date at the top of index.md, so anyone can tell which version a deck was made with.

How do you test whether an AI tool follows your brand?

Run the same prompts with and without the files and compare both results against a checklist. The difference shows what the files do. Whatever is still wrong points to a rule that is missing, vague or buried.

1. Pick three real tasks, for example a five-slide deck, a one-pager and a LinkedIn post.
2. Run each prompt in a fresh chat without the files, then with them.
3. Score both outputs against the checklist below.
4. Fix the file, not the output, and run the test again.

What to check:

• Only token colors, and no banned pairs.
• Two font families, approved weights, sizes from the scale.
• Spacing from the scale, margins respected.
• One primary button or highlight per screen or slide.
• Charts in series order, values labeled.
• Sentence-case headings, no words from the avoid list.
• No invented logos, customers, numbers or quotes.
• Assumptions stated where a rule was missing.

When the tool breaks a rule, turn the rule into a number, move it higher in the file, or add a do and don't pair. Rerun the test after every change to the files.

How is this different from llms.txt and design tokens?

llms.txt is public and tells AI assistants what your company is. Brand Markdown files are private and tell production tools how your company looks and sounds. Design tokens are the values both your code and your Markdown should come from.

llms.txt is a proposed standard: a Markdown file at the root of a site that helps AI agents use it. Ours lists our services, pages and contact details. Keep brand rules out of it. They may name unreleased products, and answer engines don't need your hex codes.

If you already have design tokens in code, generate colors.md and the type table from them instead of typing values twice. The Design Tokens Community Group released the first stable version of its format, 2025.10, in October 2025. In that format a color token is an object, not a bare hex string:

{
  "color": {
    "larkspur": {
      "$type": "color",
      "$value": { "colorSpace": "srgb", "components": [0.2, 0.275, 0.761], "hex": "#3346c2" },
      "$description": "Primary buttons, links, chart series 1"
    }
  }
}

The JSON feeds your code. The Markdown adds what a token file is not built for: when to use each value, and what never to combine.

Who should own AI brand guidelines?

One person, usually the brand or design lead. Without an owner, copies drift: one version in a Claude project, an older one in a custom GPT, a third pasted into someone's prompt.

• One owner approves every change.
• One source. The files live in the repo or one shared folder. Tools get copies from there; nobody edits a copy.
• A changelog at the bottom of index.md: version, date, what changed.
• A list of where the files are attached. Uploaded files are copies, so each one needs replacing when the source changes. Gemini is the exception for Drive files: Google says a Gem uses the most recent version of a file added from Drive.
• A review whenever the brand changes: a new color, product name or slide template. Rerun the test afterwards.

Where should you start?

Start with colors.md, typography.md and copy.md. They cover what people notice first: color, type and words. If your colors, fonts and spacing are already defined, the first version is mostly copying values into the templates above.

If your brand still lives in a PDF and your team already works with AI tools, we can turn it into a tested Markdown set, or build the brand system behind it first. See B2B brand strategy and identity, or tell us what you're working on.