Content
88%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
An unusually dense, high-signal procedural skill: concrete commands, parameters, thresholds, and validation checkpoints throughout, with almost no filler. The two weaker spots are minor over-explanation in the historical/justification passages and missed opportunities to split reference-grade detail into separate files.
Suggestions
Move the INFRA-3096/Device-Farm discussion (§ 3, "Remote devices") and the pixelmatch threshold derivation (§ 7, "Fallback thresholds") into a reference file, keeping SKILL.md to the decision rules and a pointer.
Tighten the narrative justifications (e.g. "Why two conditions, so you can adapt rather than cargo-cult") to their conclusions — the operative numbers and rules can stand without the full derivation.
Add one complete example screenshot-diff invocation (mode A, saved-PNG pair) so the primary path is fully copy-paste ready without loading the schema first.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Nearly every section carries project-specific, non-derivable knowledge (status-bar exclusion, the scale: 1.0 mandate, the Storybook/Metro mode tension) with no padding of concepts Claude already knows; however, the § 3 INFRA-3096/Device-Farm exegesis and the § 7 threshold-derivation narrative run long and could be trimmed or moved. | 4 / 5 |
Actionability | Copy-paste-ready commands (bun add -g @swmansion/argent, adb reverse tcp:8081 tcp:8081, bun nx run mobile:start:production), a concrete JSON snippet, exact parameter tables, numeric thresholds, real file paths, and a numbered five-step procedure cover the common cases fully; the only omission is a literal full screenshot-diff invocation, deliberately deferred to the loaded tool schema. | 5 / 5 |
Workflow Clarity | The § 5 procedure is a clear numbered sequence with explicit validation checkpoints (settle-before-capture, retake on transitional frames, read summary then corroborate against the diff images, no-change control-pair calibration), plus feedback loops for the fallback path; the destructive/batch cap does not apply since capture is read-only and validation is present regardless. | 5 / 5 |
Progressive Disclosure | No bundle files exist, and the body is well-sectioned with internal § cross-references and a Related Skills table for navigation; however, at ~230 lines the digressive material (the INFRA-3096 Device-Farm reading, the pixelmatch threshold derivation) could live in a reference file rather than inline in SKILL.md. | 4 / 5 |
Total | 18 / 20 Passed |