lookdev-auto is a Claude Code skill for automated visual tuning: a vision model rates rendered variants and suggests better values in a loop.
When "looks/feels right" is the only metric, make a vision/video model the judge: render a few labeled variants into ONE contact sheet — params burned onto each — ask the model to rate them and suggest better values as JSON, render the suggestions, pick the best, repeat. Usually ~2 rounds.
One round of a color-grade tune. Six variants → one upload → one inference: the judge scores each cell (the labels are burned on, so it sees param + result together) and returns {"ratings":…,"best_so_far":"D","suggest":[[warm,γ]…]}. Render the suggestions next round, ask it to pick the single best. (Scores here are an illustrative round, not a live API call.)
Claude Code is text-first, so for anything where success is "does this look or feel right" — easing, zoom feel, a color grade, spacing — it tends to guess a number and ask you to react in prose. Slow and lossy. When a real numeric metric exists, optimize that. When one doesn't, a vision model (for spatial looks) or a video-understanding model (for motion/timing) can stand in as the eye, inside a tight loop, so you stop hand-tuning by trial and error.
What this is: not a new technique — it structures an ability Claude already has (look at the output, iterate) into a disciplined, repeatable harness. The model is the eye; you do the rendering and run the loop.
- 🖼️ One artifact per round, not one call per variant — montage N variants into a single contact sheet (images) or a labeled sequence (video/motion). A 6-variant round is 1 upload + 1 inference, not 6.
- 🏷️ Params burned onto the artifact — each cell carries its own label (
A · warm −10 · γ1.2), so the judge sees label + result together; no separate "variant A used X" context to carry, fewer tokens, fewer mistakes. - 📊 Structured JSON out — the model returns per-variant ratings + concrete suggested next values (
{"ratings":{…},"best_so_far":"X","suggest":[[p1,p2]…]}); parse it, no free-text wrangling. - 🎯 Coarse → fine — round 1 is a wide, sparse spread to locate the region; round 2 renders the model's suggestions (plus the carried best) and asks it to pick the single best. Converges in ~2 rounds.
- ⚓ Calibration anchors — include one deliberately-bad and one safe-default variant each round, so the judge has a reference scale and a bad recommendation is caught (its "best" worse than the safe default = stop).
- 🧪 Independent rubric, stated up front — define "good" / "too much" / "too little" concretely instead of "which do you like", so the judge stays independent of your framing.
- ⏭️ Early-exit — if round-1 top rates ≥9/10 and the suggestions cluster, skip round 2 and apply the winner.
| Use it when | Don't — instead |
|---|---|
| "Looks/feels right" is the bar and there's no cheap numeric metric | A real metric correlates with quality → optimize that directly |
| The judgment generalizes (smooth easing, clean grade, balanced layout) | The judgment is the user's taste/brand → show them the variants, let them pick |
| ~3–6 variants worth comparing | One or two variants → just look yourself |
/plugin marketplace add connerkward/ckw-skills
/plugin install lookdev-auto@connerkward
Standalone (this repo only):
/plugin marketplace add connerkward/lookdev-auto-skill
/plugin install lookdev-auto@lookdev-auto
Or drop this repo's SKILL.md into your agent's skills directory.
🗂️ Part of ckw-skills
The automated-eye sibling of lookdev (a human dials it in by eye) and deterministic-design (measure the UI instead of trusting any eye). Here a vision model is the eye.
MIT © Conner K Ward
🧭 ckw-skills — part of Conner K. Ward's collection of Claude Code skills & MCP servers.