self-improve

ATELIER — Self-improve

The design half of the marketplace's self-improving loop. A design that passed the fitness rubric yet still failed a real user — a hierarchy that confused, a contrast that broke on a device, a flow that dead-ended — is the signal: the fix is a better-articulated canon rule or a re-weighted rubric dimension, fixed once, upstream — not a louder reviewer. (Covenant: ../../knowledge/covenant.md.)

Feedback intake (cross-project learning)

ATELIER gets sharper from design feedback — from a downstream user, from a real-usage signal, or from its own loop stalling. Each entry is symptom → which canon/rubric gap let it through → what rule would have caught it (e.g. "the CTA tested fine but users missed it → the hierarchy dimension didn't penalise a primary action that loses the Von Restorff contrast; add that check"). A loop that stalls below target (diminishing-returns halt) is itself feedback: the reviewer couldn't name a converging fix.

The loop

1. Reflect on the target element (a canon lens, a rubric dimension/weight, the loop's stop-conditions, a skill) against the covenant + pillars: is a canon rule missing a forcing check? does a rubric weight under-value a real failure? does the loop ping-pong (→ tighten the delta-floor / fix the reviewer)? does a skill do more than one thing (→ cleave it)?
2. Decide: sharpen a canon rule (../../knowledge/canon/README.md) / re-weight or add a rubric dimension (../../knowledge/protocols/design-critique-loop.md) / cleave / reference-not-restate. If the element is already tight, say so and stop.
3. Apply on a branch (surgical, one concern).
4. Adversarial gate: run /foundry:pr-review if foundry is installed; otherwise self-review against the covenant and state the gate ran in reduced form.
5. Open a PR for the human to merge (never self-merge), following the user's merge governance (FOUNDRY's pr-approval default), so the improvement reaches every user. When foundry is present, offer the same lesson to its source-level design-critic so the discipline compounds across the build.

Each pass must leave ATELIER measurably better at catching real design failures before they ship — at least halving the remaining distance. Record the recurring failure-class so the next like design is caught by default.