takazudo/design-polish-loop

Design Polish Loop

Level up the look of a page that already renders. Each round looks at the actual pixels, finds the gap to a better-looking target, and closes it with on-system CSS — then re-captures and checks. The target is proposed by image generation, so the loop reaches for a genuinely nicer design rather than orbiting the current one.

This is a thin orchestrator — it owns the loop, not the techniques. It drives skills that already exist; its value is wiring them into a capture → propose → diagnose → polish → re-capture cycle with the right discipline at each step.

| Step | Skill it drives | Role | | --- | --- | --- | | Capture | /headless-browser | screenshot the live page at real breakpoints, worst-case content | | Propose direction | /codex-imagegen | turn the current screenshot into polished "north-star" mockups (ChatGPT-billed) | | Judge & diagnose | /design-iteration-wisdom | render→look→diagnose from pixels; reference-match against the chosen direction | | Constrain the fix | /css-wisdom | polish using the tight-token strategy — system tokens, never arbitrary px/hex | | Measure | /verify-ui | settle color / exact-value questions the eye can't (sample, don't eyeball) |

The novel ingredient is imagegen as the direction engine. A model handed the existing source over-anchors on it and "polish" collapses back to the original. A north-star mockup generated from the screenshot breaks that anchor — it shows a better look the loop then implements with real tokens. The mockup is direction, not spec (see caveats): mood, hierarchy, spacing rhythm, color, decorative treatment — never its exact text/numbers/components, which it hallucinates.

Input parsing

• count (number): polish rounds. Default 2.
• route-or-url: the page to polish (e.g. /pricing or http://localhost:5173/pricing). If omitted, ask which page, or infer the single obvious WIP route.
• -n <directions>: how many north-star mockups to generate up front. Default 2. Each costs real ChatGPT usage — see Cost.
• --no-imagegen: skip mockups entirely; run a pure /design-iteration-wisdom self-judged polish (free, but only nudges the current look — no fresh direction).
• --target <image>: use a supplied reference image as the north-star instead of generating one (skips imagegen; the reference-match discipline still applies).
• --cleanroom: implement the polish in an isolated prototype (/prototype-first-wisdom) instead of in-place. Use when the chosen direction is a big departure that would fight the existing markup; otherwise default in-place polish on the WIP.

Workflow

Step 0 — Detect context

Per /design-iteration-wisdom Step 0: find the design-system / token source (a project *design-system* skill/doc, or the existing token scale as de-facto), the dev-serve command + port (reuse a running server, don't spawn a duplicate), and the target route(s). Load /css-wisdom tight-token articles now so polish stays on-system from round 1 — at minimum methodology/design-systems/tight-token-strategy/index.mdx and component-tokens.mdx (the system-tokens-vs-arbitrary-values decision framework), plus color-tokens.mdx / typography-tokens.mdx / token-preview.mdx as the change demands.

Step 1 — Capture the baseline

Serve the page, then /headless-browser screenshot at ~390 / ~760 / ~1300px, on the worst-case content (longest title, longest list, near-empty group), not the page top. Read every PNG. This baseline is both the imagegen input and the before-shot.

Step 2 — Propose north-star directions (unless --no-imagegen / --target)

Feed the baseline (usually the desktop shot) to /codex-imagegen once per direction:

$HOME/.claude/skills/codex-imagegen/scripts/codex-imagegen.sh \
  --in <baseline-desktop.png> --out direction-1.png \
  --prompt "Polished, premium redesign of this screen. Keep the same content, structure, and information. Elevate spacing rhythm, typographic hierarchy, color, depth, and detailing. Modern, restrained, on-brand — not generic AI gradient slop."

Vary the brief per direction (e.g. "editorial / high-contrast", "calm / airy", "denser / utilitarian") to spread the options. Read the mockups, show them to the user, and let the user pick the direction — this is a taste call only a human owns. The picked mockup is the round target.

Step 3 — Polish rounds (repeat count times)

Each round is /design-iteration-wisdom with the chosen direction as the reference:

1. Diff, don't admire. Crop the current render to the mockup's framing and walk it element-by-element against the direction (reference-match discipline). Name what's behind: grouping/contrast, hierarchy, type measure, color, weight, depth. Because an imagegen reference is a fuzzy spec, match its design intent (rhythm, hierarchy, palette, treatment) — do not reproduce its invented copy, fake rows, or un-buildable flourishes.
2. Diagnose in token terms. Map each gap to a move on the project's scale, using the /css-wisdom tight-token framework. "Cards read flat → raise elevation one step + tighten in-card gaps, widen between-card gap" beats "add shadow". Never introduce an arbitrary value where a token exists.
3. Apply the change with project tokens/components, minimal and targeted.
4. Re-capture at all breakpoints; measure anything near-threshold with /verify-ui (color especially — the eye lies; sample computed styles).
5. Early exit when a round closes no meaningful gap, or the render reads cleanly against the direction at every width. Diminishing returns end the loop before count.

Commit each accepted round (or stash before a risky one) so any round is revertible.

Step 4 — Stop and hand to the human

Design has a taste component an AI approximates but doesn't own. When the computable invariants pass (/design-iteration-wisdom → contrast ≳2×, monotonic hierarchy) and it reads clean, stop and show before / after / north-star side by side for the final call. Don't loop chasing a verdict only a person can give.

Cost

Each north-star mockup is one /codex-imagegen call — roughly 90k–230k ChatGPT tokens (photoreal/complex prompts cost more). Generate directions once up front, not per round; the rounds themselves are free (capture + CSS). Re-invoke imagegen only to explore a new direction or when the loop plateaus and a fresh target is wanted. --no-imagegen removes the ChatGPT cost entirely.

Important notes

• Imagegen is direction, not spec. Raster mockups can't be traced to code and hallucinate text/numbers/components. Use them for look and feel; take the real content from the live page. This is why the implementation is governed by tight tokens + the existing design system, not by the PNG.
• Polish stays on-system. The whole point of binding to /css-wisdom's tight-token strategy is that improvements move along the project's scale. An arbitrary #hex or 13px to "match the mockup" is a regression even if it looks closer — flag it, find the token.
• In-place by default; clean-room for departures. Incremental polish edits the WIP in place. A big directional jump that would fight the markup belongs in a /prototype-first-wisdom clean-room first (--cleanroom), then ported back deliberately.
• Expect trial and error. Directions that don't land get discarded; some rounds regress and get reverted. The loop is exploratory — the value is the design you converge on, and the human picks direction and final acceptance.
• Two human checkpoints: picking the direction (Step 2) and accepting the result (Step 4). Everything between is mechanical capture-diagnose-polish.