Chat is the wrong UI for design judgment. A piece on UX Design argues that designers should stop trying to talk AI coding agents through decisions in real time — and instead bottle that judgment into a structured DESIGN.md file the agent can read.
The core claim: AI doesn’t just need values. It needs the reasoning behind them. A hex code without context produces generic output. A hex code paired with “this is the primary accent, it should feel clear and trustworthy” gives the agent something to optimize against.
What a real DESIGN.md covers:
- Color palette — exact hex values plus intent (“primary accent for trust signals”)
- Typography — scale ratios, weights, and where each style is used in the UI
- Components — naming conventions, spacing rules, state definitions (hover, focus, disabled)
- Layouts — grid systems, breakpoints, responsive behavior per viewport
- Accessibility — contrast minimums, focus states, semantic HTML requirements
The failure mode is predictable. Teams create a DESIGN.md, connect it to Claude Code or another coding agent, and the output still looks generic. Correct brand color, wrong everything else. Components stay inconsistent. Mobile behavior breaks. Accessibility issues persist. The root cause isn’t the AI — the file contains values without judgment.
NOTE
The article frames this as “bottling judgment” — extracting what experienced designers know but can’t always articulate into a format the machine can act on. The AI doesn’t replace the decision. It forces you to make the decision explicit.
The practical shift: stop debating design choices in chat. Write them down with reasoning, hand the file to your coding agent, and let it work against a fixed spec instead of guessing from conversation fragments.
DESIGN.md works when it carries both the “what” and the “why” — values alone produce AI slop with the right brand color.