compare-screenshotslisted

Compare two screenshots and report every visual difference between them. Use this skill whenever the user gives you two image files (before/after, expected/actual, v1/v2, baseline/candidate) and wants to know what changed, spot the differences, diff two UI states, do visual-regression checking, or confirm two screens look identical. Triggers on "compare these screenshots", "what's different between these two images", "diff these UIs", "did the layout change", "spot the difference", "/compare-screenshots", or any before/after image pair. The skill first checks the pair is similar enough to be worth comparing, then produces a similarity percentage, a third image with the changed regions boxed in red, and a written rundown of what actually changed in each region.
Stoica-Mihai/claude-skills · ★ 0 · AI & Automation · score 70

Install: claude install-skill Stoica-Mihai/claude-skills

# Compare Screenshots Two screenshots in, three things out: a **similarity score**, a **highlighted overlay image** with each changed region boxed and numbered, and a **written rundown** of what changed in each box. The work splits cleanly: - **A bundled script does the pixel math** — alignment, similarity scoring, the similarity gate, region clustering, and drawing the overlay. This is deterministic and exact; eyeballing pixels is not, so do not try to compute diffs by hand. - **You do the semantic part** — looking at the boxed regions and the cropped before/after pairs the script saves, then describing *what* changed ("the button turned green", "the heading reads 'Hello there' instead of 'Hello world'"). The script knows *where* differs; only vision knows *what* it means. ## Why the similarity gate exists If the two images are of completely different screens, a pixel diff is noise — every region "changed", and a rundown would be meaningless. So the script scores structural similarity (SSIM) first and refuses to diff a pair that falls below a threshold, telling the user the images are too dissimilar to compare. This keeps the output honest: a diff is only produced when there is a shared structure to diff against. ## Workflow 1. **Resolve the two image paths.** The user supplies them (often as command arguments). If only one or zero paths are given, ask for the missing one — do not guess. 2. **Pick an output directory** for the overlay and crops. Def