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 exemplary instruction-only skill body: concrete tool-call recipes for the common cases, a strict survey-first workflow with explicit verification after every action, real error codes with recovery paths, and genuine safety stops for sensitive apps. The only room for improvement is minor trimming of repeated warnings and moving the sample Survey output / error catalog into a reference file to slim the main body.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and tool-specific rather than padded — nearly every section teaches behavior Claude cannot know ("Cached per session × app for 5 minutes", "covered / off-screen / minimized windows still capture", "richness: 0.995"). Minor over-explanation remains: the "When a human opens an unfamiliar app they look around first" analogy, and the self-drawn/close-button warning repeated across three sections (Core mindset, Execution priority, and Rules 1/4), which could be trimmed. This fits the score-4 anchor (efficient, minor instances that could be trimmed) rather than score 5, whose 'every token earns its place' bar the repetition and analogy fall short of; it is well above score 3's 'some unnecessary explanation'. | 4 / 5 |
Actionability | Flows A–F give concrete, directly issuable tool invocations covering the common cases — e.g. "MacosAct { type: 'intent', app: 'WeChat', intent: 'open_moments' }", "MacosAct { type: 'key', keys: ['cmd', 'n'] }", "MacosCaptureWindow { windowID: 227474 }" — with real parameter shapes, concrete error codes (`WINDOW_CHROME_BLOCKED`, `permissionsMissing`) and their remedies. Placeholders like `path: [...]` are inherently runtime-determined values, not pseudocode. This matches the score-5 anchor (specific examples cover the common cases); it is above score 4 because no real gaps exist beyond values only knowable at runtime. | 5 / 5 |
Workflow Clarity | A clear three-phase sequence ("Survey → Plan → Act+Verify"), a strict execution priority order ("Go down the list in order. Never skip ahead."), and explicit validation checkpoints: "**Observe after every MacosAct.** UI has state; you can't predict the outcome", plus feedback loops for error recovery ("Click has no effect → re-observe", "WINDOW_CHROME_BLOCKED → … Pick one of the 4 suggested fallbacks") and stop conditions ("After two coord clicks with no visible change, stop"). This matches the score-5 anchor (explicit validation steps, feedback loops, checklists via the Rules section); score 4's 'minor validation gaps' do not apply. | 5 / 5 |
Progressive Disclosure | The single SKILL.md (~180 lines) is well-sectioned with headers, a tool inventory table, and labeled flows, and no bundle files exist to mis-link. However, all content is inline with zero one-level-deep references — the "What a Survey response looks like" annotated sample and the Error diagnosis catalog are candidates for a reference file — so it fits the score-4 anchor (good structure, most content appropriately placed, minor organization gaps) rather than the score-5 anchor, which requires well-signaled external references; the under-50-line exception for a reference-free skill does not apply at this length. | 4 / 5 |
Total | 18 / 20 Passed |