VercelEngineering PostOfficial Source

How our agents build on-brand pages with design.md

Vercel found that prose alone produced inconsistent agent-made pages, then paired design.md with fixed CSS primitives and repeatable evals to encode brand judgment.

Vercel · Aug 31, 2026

Open Source Open Markdown Open JSON

Source Summary

Vercel built **design.md** as a public guidance file, paired it with a public stylesheet and an evaluation loop. It tests **seven fixed scenarios** across Claude Opus 4.8 and Codex with GPT-5.5.

Practical Implication

For agent-built interfaces, encode reader intent and composition rules in context, move mechanical choices into reusable CSS, and rerun frozen prompts after every guidance change. Keep human review for hierarchy and deterministic checks for layout failures.

Agent-Ready Context

Vercel built **design.md** as a public guidance file, paired it with a public stylesheet and an evaluation loop. It tests **seven fixed scenarios** across Claude Opus 4.8 and Codex with GPT-5.5.

For agent-built interfaces, encode reader intent and composition rules in context, move mechanical choices into reusable CSS, and rerun frozen prompts after every guidance change. Keep human review for hierarchy and deterministic checks for layout failures.

The reported comparison used one generation per condition with no rerolls. It shows that guidance changed one page's structure, but does not establish general reliability across models, prompts, or brands.

Connected Context · Feed7 Judgment

This makes shared design guidance more concrete: semantic composition rules belong in agent-readable context, mechanical decisions in reusable CSS, and changes behind frozen prompts plus human and deterministic review. It supports harness engineering over ad hoc prompting, while the seven-scenario, single-sample evaluation sharply limits any claim that the method generalizes across brands, prompts, or models.

Agents, codebases, and teams — Aditya Khandelwal, Amazon AGI Labdesign.md is a concrete instance of maintained team context paired with automated checking, reinforcing the view that agent quality depends on shared repository infrastructure rather than individual prompts.How we set up our cloud agent environmentCursor’s environment work supplies a prerequisite: design guidance and reusable CSS are useful only when agents can reliably discover, run, and verify the surrounding workflow.addyosmani/agent-skillsThe skill pack’s evidence gates align with frozen prompts, deterministic layout checks, and human hierarchy review, turning guidance changes into reviewable workflow changes rather than prompt intuition.The Era of Compound Engineering — Kieran Klaassen, Every/CoraThe evaluation loop operationalizes compound engineering by preserving reviewed design judgments and testing whether revised context improves later runs; its fixed scenarios also impose the curation discipline that accumulated memory needs.

Context Map

agentcoding#skills#harness-engineering#design-engineering

Uncertainty

The reported comparison used one generation per condition with no rerolls. It shows that guidance changed one page's structure, but does not establish general reliability across models, prompts, or brands.