# Does a DESIGN.md keep your brand? We measured the drift

Updated 2026-11-21 · https://briefatto.com/guides/does-design-md-keep-your-brand-drift-test/

Yes: any DESIGN.md does, and no file does not. In an empty repo where the brand could travel only by file, Claude Code and Codex matched the brand's fonts in 0 of 18 pieces without a design file and in 53 of 54 with one, and their colours stopped drifting between sessions.

## The result

- **0 of 18**: pieces matched the brand's fonts with no design file.
- **53 of 54**: pieces matched them with a DESIGN.md, in any of three formats.
- **0.40 or less**: session-to-session palette drift (ΔE2000) with any DESIGN.md, against 1.7 to 12.3 without one.

Without a file, the agent invents a brand every session. The palette moved by 1.7 to 12.3 CIEDE2000 between sessions, and as little as 31% of a piece's paint came from the brand's colours. With a DESIGN.md the drift fell to 0.40 or less, below the 1.0 where a difference starts to be visible, and brand colours covered 99% of every piece. Without one, 12.3 is a different colour.

| Setup | Brand fonts matched | Brand colour coverage | Session-to-session drift (ΔE2000) |
| --- | --- | --- | --- |
| A. No design file | 0 of 18 | 31% to 97% | 1.7 to 12.3 |
| B. awesome-design-md style | 18 of 18 | 99% | 0.02 to 0.23 |
| C. Google design.md spec | 17 of 18 | 99% | 0.12 to 0.40 |
| D. Briefatto Board | 18 of 18 | 99% | 0.05 to 0.31 |

Raw numbers per piece: [drift-test-results.csv](https://briefatto.com/assets/drift-test/drift-test-results.csv).

## Every session, side by side

Each row is one setup; each column is a fresh session that never saw the others. Look at row A, then at the three rows below it.

**Social card** Row A: no design file. B: awesome-design-md style. C: Google design.md spec. D: Briefatto Board.

- Codex, three fresh sessions per row, [picture](https://briefatto.com/assets/drift-test/design-md-drift-test-social-card-codex-three-sessions.png)
- Claude Code, three fresh sessions per row, [picture](https://briefatto.com/assets/drift-test/design-md-drift-test-social-card-claude-three-sessions.png)

**Landing-page section** Row A: no design file. B: awesome-design-md style. C: Google design.md spec. D: Briefatto Board.

- Codex, three fresh sessions per row, [picture](https://briefatto.com/assets/drift-test/design-md-drift-test-landing-section-codex-three-sessions.png)
- Claude Code, three fresh sessions per row, [picture](https://briefatto.com/assets/drift-test/design-md-drift-test-landing-section-claude-three-sessions.png)

**Instagram post** Row A: no design file. B: awesome-design-md style. C: Google design.md spec. D: Briefatto Board.

- Codex, three fresh sessions per row, [picture](https://briefatto.com/assets/drift-test/design-md-drift-test-instagram-post-codex-three-sessions.png)
- Claude Code, three fresh sessions per row, [picture](https://briefatto.com/assets/drift-test/design-md-drift-test-instagram-post-claude-three-sessions.png)

## The file is the protocol

A coding agent starts every session from zero. Give it a brief and no design file, and it draws a good page and a different brand each time: a dark green card in the first session, a violet one in the second, an orange one in the third. Row A in every sheet is that slot machine.

An agent has no memory between sessions; the design file is its memory. That makes the file a protocol, not a style guide: once it is in the repo, Claude Code and Codex land on the same fonts and the same colours, session after session. The format mattered less than having the file.

Writing the file is the easy part. Keeping it true is the job. Briefatto Board builds it from your live site with every colour, font and radius measured from a real render, stamps it with a version and checks every piece the agent renders against it. When the brand changes, `board status` names every piece that is now stale.

## Method

|  | What was fixed |
| --- | --- |
| Brand | Issue Composer, an MIT open-source tool, pinned at commit 151bb54; its live page is the ground truth for colours and fonts |
| Setting | Every session started in a new, empty git repository holding only the setup's design file: no site source, no screenshots, no web access to the site |
| Setups | A: no file. B: a DESIGN.md in the awesome-design-md style. C: a DESIGN.md to Google's design.md spec, lint-clean. D: the DESIGN.md and skill Briefatto Board produced from the site |
| Agents | Codex on gpt-6.2-sol and Claude Code on Claude Sonnet 5.5, three fresh sessions per setup each, three requests per session: 18 pieces per setup |
| Requests | A 1200x630 social card, a 1440x900 landing-page section and a 1080x1080 Instagram post for one invented release |
| Scoring | By program, on the final renders: share of painted pixels within ΔE2000 5 of the site's palette, computed font families against the site's, and mean pairwise palette distance between sessions. No judges |
| Date | 2026-11-21 |

## Questions and answers

### Does Claude Code keep my brand without a DESIGN.md?

Not across sessions. With no design file, Claude Code and Codex matched the brand's fonts in 0 of 18 pieces and their palette drifted 1.7 to 12.3 ΔE2000 from one session to the next.

### Which DESIGN.md format should I use?

Any of the three. All held the fonts in 17 or 18 of 18 pieces and kept drift at 0.40 or less; having the file is what counts. Briefatto Board writes one from your site with measured values, versions it and checks every rendered piece against it.

### How do I get a DESIGN.md for my site?

Run `npx briefatto-board install`, then ask your agent to adopt your site. Board captures the pages, the agent writes the guide, and `board freeze` measures and stamps it as version one.

### Can I check the numbers?

Yes. The site is public and pinned, the per-piece numbers are in the CSV, and every session's pieces are in the sheets on this page.

## Try it

```
npx briefatto-board install
```

Briefatto Board is free and open source under Apache-2.0: https://github.com/Briefatto/briefatto-board

## Related guides

- [Briefatto Board vs Impeccable, Taste Skill and Hallmark: the blind test](https://briefatto.com/guides/impeccable-vs-taste-skill-vs-hallmark-blind-test/)
- [The best DESIGN.md generators and libraries in 2026](https://briefatto.com/guides/best-design-md-generators-2026/)
- [The best Claude Code design skills in 2026](https://briefatto.com/guides/best-claude-code-design-skills-2026/)
- [What is DESIGN.md, and how do coding agents use it?](https://briefatto.com/guides/what-is-design-md/)

## Sources

- https://en.wikipedia.org/wiki/Color_difference#CIEDE2000 (checked 2026-11-21)
- https://github.com/Briefatto/briefatto-board/blob/main/README.md
- https://github.com/Briefatto/briefatto-board/blob/main/LICENSE
