Design Eval Loop

Drive an existing UI artifact to a verified 5/5 on a current design rubric. The engine is a per-view loop: enumerate → audit + score (parallel sub-agents) → fix → re-score, repeated until every view passes, then once more for the whole app. It is a finishing pass — the artifact must already exist.

This skill exists to stop two failure modes: shipping dead controls (links and buttons that go nowhere) and claiming a false 5/5 without evidence.

EC public-safety note

This skill is safe to publish as a generic workflow. Do not include private product names, customer data, internal screenshots, private URLs, or local filesystem paths in public scorecards. For public release, run the publish-skill sanitizer before sending the bundle to GitHub, SuperAda, or ClawHub.

When to use

• "Score this and keep improving until it's 5/5 on the design eval."
• "Make sure every page is linked — no dead items."
• "Run sub-agents over every view and rate them till they pass."

Not for: building the artifact from scratch (build it first), or a one-line visual tweak (just make the edit).

Inputs to lock before looping

1. Target file(s) — which artifact(s). If multiple builds exist, confirm whether the loop runs on one or all of them; they are separate files and each needs its own pass.
2. View inventory — the full list of distinct views/screens/panels/overlays. Derive it from the file, don't guess. State the count aloud (e.g. "20 views").
3. Rubric — default to references/rubric.md. Honor any dimension the user names explicitly (e.g. "2026 best practices", "WCAG AA").
4. Pass bar — default 5.0/5 per view, then 5.0/5 whole-app. The user may relax it (e.g. "≥4.5 is fine").

The loop

Plan it with TodoWrite first — one todo per phase, plus the iteration cap.

Phase 1 — Enumerate

List every view and every interactive element. A view is any distinct surface: home, drawer, sheet, overlay, settings list, and each detail panel. Detail panels are views too — don't fold them into one line.

Phase 2 — Audit + score in parallel

For each round, spawn two sub-agents at once (one message, two Agent tool calls) so they don't collide on the file:

• Agent A — Dead-item audit. Crawl every clickable element; report any that navigate nowhere, have no handler, or give no feedback. Output: a list of dead items with selector + expected destination.
• Agent B — Design eval. Score every view against references/rubric.md, 0–5 per dimension, with a one-line reason per sub-score and the specific blocker keeping each view under the bar.

Use the prompt templates in references/sub-agent-prompts.md verbatim; they force structured, file-grounded output instead of vibes.

Phase 3 — Fix

Apply the highest-impact findings first: real bugs (undefined tokens, occluded panels, wrong z-index), then dead items, then the lowest-scoring rubric dimensions. Fix in the canonical file. After edits, run a static integrity check (balanced tags/braces/parens, every panel defined, every row mapped).

Phase 4 — Re-score

Re-run Agent B (a fresh agent — don't trust the fixer's self-report) on the changed views. Record the delta in a scorecard. Repeat Phases 2–4 per view until it hits the bar.

Phase 5 — Whole-app pass

Once every view passes individually, score the whole app as one experience (cross-view consistency, shared theme, navigation coherence, no orphan surfaces). Iterate until the whole-app score hits the bar.

Convergence and escape hatches

• Iteration cap: max 4 rounds per view and 3 whole-app rounds. If a view won't reach the bar within the cap, stop and report the exact remaining blocker — do not loop on cosmetics.
• Honesty gate (non-negotiable): never claim 5/5 without a fresh scoring agent confirming it. If the real ceiling is structural (e.g. native-control rewrite, focus-trap) and out of scope, say so and report the true score rather than rounding up. A reported 4.6 with a named gap beats a fabricated 5.
• User break: the user can stop the loop at any round; report current state.

Output each round

Emit a compact scorecard table (view → before → after) plus an <od-card type="verify-scorecard"> covering the dead-item and rubric checks. End with: which views pass, which don't, the named blocker for any that don't, and the next round's target.

Hard rules

• Parallel, not serial: Agent A and Agent B launch in the same message.
• Re-score with a fresh agent: the agent that fixed the code does not grade its own work.
• Every link opens somewhere: zero dead items is a P0, not a nice-to-have.
• Detail panels count as views. A 5/5 on the shell with unscored panels is not a 5/5.
• One canonical file is the source of truth — edits land there, not in chat.

References

• references/rubric.md — the scoring dimensions, weights, and 5/5 bar.
• references/sub-agent-prompts.md — paste-ready prompts for Agent A and Agent B.