The agents' own docs say keep it short
Claude Code's documentation calls CLAUDE.md context, not enforced configuration, and says the more specific and concise the instructions, the more consistently Claude follows them. Anything that must happen at a fixed point belongs in a hook. Cursor's documentation says to keep rules under 500 lines and to add one only when the Agent repeats a mistake.
A list of "never use purple gradients" rules grows with every bad screen, and it still leaves the next case open.
Give each rule to whoever can enforce it
- Measurable rules go to code. Clipped text, text off the canvas or over other text, type too small where it is seen, fonts that never loaded, contrast below WCAG and the Impeccable detector's anti-patterns are already in
board check, which fails the piece with exit status 1. - Brand facts go to the record. Freeze writes the exact computed styles into DESIGN.md. Each confirmed must or must-not goes in its own fenced
constraintblock inrefs/notes.md, and freeze copies them into the guide. - Taste goes to the model. Briefatto Board ships no taste rubric. The model makes the aesthetic calls, and the intent reaches it as pictures: a reference sheet and three drawn directions you choose from.
What stays in the instruction file
One block. After a freeze, the Briefatto Board block in AGENTS.md and CLAUDE.md says three things: read DESIGN.md before any visual work, route brand changes through board freeze, and run board check on every piece before handing it over.
Questions and answers
Should I delete the design rules I have?
Move them. Check the provable ones against what board check already fails, put the brand musts and must-nots in constraint blocks, and drop the adjectives once the guide is frozen.
Doesn't a long rule list at least help a little?
Claude Code's documentation says longer files consume more context and reduce adherence. A short pointer to a frozen guide carries the exact values in a few lines.