Guides · Study · November 5, 2026
Given the same brief, Claude’s pages scored 1.7 points higher out of 10 for brand fidelity with a full brand guide and skill than with a DESIGN.md file alone (9.0 vs 7.3; 6.6 with only the brand’s name), in a blind grade of 12 brands.
The full guide plus skill beat DESIGN.md for 12 of 12 brands, by 1.7 points on average.
Its best dimensions were color (8.3) and typography (8.3). On imagery it scored 5.3, the same as the brand name alone (5.3). The guide scored 9.1.
DESIGN.md added 0.7 points over the brand name alone, and scored lower for 4 brands: Shopify, IBM, Netflix, Spotify.
| Dimension | Name only | DESIGN.md | Guide + skill | Guide gain |
|---|---|---|---|---|
| Color | 7.7 | 8.3 | 9.6 | +1.3 |
| Typography | 6.7 | 8.3 | 8.9 | +0.6 |
| Layout | 6.3 | 7.4 | 8.9 | +1.5 |
| Components | 6.9 | 7.7 | 8.9 | +1.2 |
| Imagery | 5.3 | 5.3 | 9.1 | +3.8 |
| Signature moves | 5.8 | 6.5 | 9.0 | +2.5 |
| Voice | 7.4 | 7.9 | 8.5 | +0.5 |
| Recognition | 6.3 | 6.8 | 9.2 | +2.4 |
A landing page for a one-day customer conference, “[Brand] Live”: nav, hero, three benefits, six sessions, a sign-up form, footer. The same text for every brand.
Read the briefA: the brand name only. B: the brand’s DESIGN.md only. C: the full brand guide from our study plus its Claude skill and tokens.css. Each builder saw only its own input, with web access off.
The DESIGN.md files and skillsClaude Opus 5.5 (claude-opus-5-5), as Claude Code subagents built each page in one pass, with no screenshots and no fixes. 36 pages in all.
A different model (OpenAI gpt-6.1-sol through the Codex CLI 0.165.0, reasoning effort high) saw the brand’s real homepage and the three pages as X, Y and Z in a shuffled order, and scored each on 8 dimensions from 1 to 10. Every brand was graded twice with new orders.
Read the rubricInternal blind grading, not peer reviewed.
The 1 to 10 scale is not calibrated. In an earlier internal test, this style of rubric scored brands’ own published work no higher than AI builds. A high score here means close resemblance, not work that passes as the brand’s own.
The grader scored the same page 0.3 points apart on average between its two runs.
12 brands, one page type, one build per brand and input. The 95% ranges above are from resampling brands.
The guide holds the brand’s real image and logo URLs. DESIGN.md can’t carry images. Part of the gap is images, and that is the point of the comparison.
We wrote the DESIGN.md files and the guides from the same studies. A DESIGN.md written another way may score differently.
The model already knows these brands, which helps the name-only pages. A less known brand would likely score lower without a file.
Graded at 1440 pixels wide, from the fold and the two screens below it.
Free to reuse with credit: “Unfairly brand fidelity benchmark, November 5, 2026”, linking to this page.
Made from your website, free.
AI brand export56 DESIGN.md files to copy, and the spec explained.
Awesome DESIGN.mdRules in a skill, visuals in a design system.
Brand guidelines in ClaudeA senior design lead and an engine make the work in your brand.
AI design agencyNot on its own in this test. Pages built from a DESIGN.md averaged 7.3 out of 10 against 9.0 for a full brand guide plus skill, on the same brief and the same grader. DESIGN.md did best on color and type, the parts it can write down as tokens.
The biggest gains were imagery and signature moves. A guide carries the brand’s real images, logo files and layout rules. DESIGN.md is a token file and can’t carry those.
A different AI model (OpenAI’s, through the Codex CLI) saw each brand’s real homepage and the three pages as unlabelled candidates in a shuffled order, then scored each on 8 dimensions from 1 to 10. Each brand was graded twice with different orders and the scores averaged.
No. It is an internal benchmark with a model grader. We publish the brief, the rubric, every score and the grader’s notes so anyone can check it or run it again.
12 brands from our public brand studies: Airbnb, Duolingo, Linear, Spotify, Slack, GitHub, Netflix, Headspace, Monzo, IBM, Shopify, Starbucks.
Yes. All 56 DESIGN.md files and Claude skills are free on GitHub and on the DESIGN.md guide. The free AI brand export makes both from your own website.