Scope: one screen, or one tight flow (2-3 screens sharing a job). More than that is mobile-redesign-app.
Method files, read in full first: ../_shared/references/redesign-protocol.md (audit,
levers, fix loop, risk governor) and ../_shared/references/review-rubric.md (finding
format). ../_shared/references/tells.md supplies the findings vocabulary.
Flow
- Stack sniff + audit (redesign-protocol.md): the screen's current tokens or inline styles, its dial reading, its place in the route tree, signature interactions, accessibility wins. Screenshot the BEFORE state if a device/preview is reachable.
- Mode. Preserve vs overhaul - ask once if ambiguous.
- Findings for this screen only, rated high/medium/polish.
- Propose the pass. Which levers (type/spacing → color-to-tokens → states → feedback/motion → recomposition), which findings each closes, what will NOT change. One short message; proceed on approval or if the user already said "just do it".
- Fix loop (redesign-protocol.md): one finding per atomic commit, before/after pair per visual change, risk governor (stop > 20%, cap 20 fixes).
- Close out. Findings closed vs remaining, the before/after pairs, and - if
design-baseline.jsonorMOBILE-DESIGN.mdexist - refresh them (deltas noted).
Hard boundaries
- Never changes: route name, deep link, position in the nav tree, form field semantics, learned gestures. If a finding requires one of those, report it as OUT-OF-SCOPE with a pointer to mobile-redesign-app.
- Blocks-first: if the screen is an archetype (settings, feed, auth, paywall, empty
state), diff it against the matching
../_shared/blocks/block and prefer converging toward the block over inventing a third pattern. - No scope creep: components shared with other screens get changed only when the change is invisible elsewhere or the user approves the blast radius; the risk governor's +5%/file makes this expensive on purpose.
- The screen must still pass the mobile-taste Section 14 pre-flight when done - a redesigned screen that fails pre-flight is not done.