duet-agent: Skill for Claude Code

Compare Screenshots

Decide which image is less wrong against what the scene should show — not whether the candidate matches the baseline. The baseline is just an earlier attempt; it can be wrong too. Treat both images as candidates measured against a target you establish yourself. Metrics locate where the images differ; they never decide who is right.

Workflow

1. Establish the target from first principles. Before looking at distance, decide what this image should show: the visual requirement, the design intent, what the thing depicts in reality, and any domain skill that owns the look. This — not the baseline — is ground truth. Write it down in one or two concrete sentences ("low sun should cast long shadows east; trees fill the canopy; labels stay legible at this zoom").
   • If the right answer isn't clear — competing valid readings, a taste or product-intent call, a tradeoff only the owner can settle — stop and ask the user what the correct answer should be. Show them the comparison. Do not quietly default to the baseline to avoid asking; that bakes in whatever the baseline got wrong.
2. Confirm comparability so the differences you see are real, not capture artifacts: same viewport, DPR, route/page, frozen time/tick, camera intent, UI state, data, fonts/assets where they matter. If not comparable, fix capture setup or compare only a crop/feature where the mismatch is harmless.
3. Generate artifacts to locate divergence, sized to the question: side-by-side, key-feature crops/zooms, grayscale, absolute grayscale heatmap, pixelmatch diff, per-image Sobel/edge maps, edge-difference heatmap, JSON metrics.
4. Judge each divergence against the target. For every place the two images differ, name what is actually there in plain terms — missing content, wrong camera, bad hierarchy, weak contrast, wrong depth, text overlap, layout shift, clipped edge, unexpected blur, style mismatch — and decide which side is closer to correct. The answer can be the candidate, the baseline, both wrong, or a genuine toss-up.
5. Get a neutral second opinion for disputed or high-stakes calls: a fresh subagent given only the two images and neutral labels, per references/subagent-visual-review.md.
6. Conclude with one verdict: candidate is less wrong (accept, and re-bless the baseline if one exists), baseline is less wrong (reject), both wrong (another pass needed — say what's still off), or unclear (ask the user). Never accept on a lower score alone or reject on a higher one. Never hide content, blur detail, crop away differences, or make the capture less truthful to move a number.