Content
92%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.
A tight, high-signal skill body: it encodes genuinely non-obvious operating policy (user-owned install, no pre-probing, explicit platform targeting, graceful failure handling) as executable commands with no token waste. The only soft spot is that interaction command syntax (tap/type/scroll) is delegated to the live snapshot hints and help topics rather than shown, which is a reasonable design but leaves the most common interaction loop one step removed from copy-paste-ready.
Suggestions
Add one copy-paste-ready example of a basic interaction verb (e.g., a tap or type command against the running app) so the most common loop — open, then act — is fully executable from SKILL.md without consulting help.
Consider a one-line pointer that the interactive snapshot itself lists available next-step commands, so readers know where interaction syntax comes from before hitting the help topics.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean with zero padding: it never explains what an iOS Simulator or a simulator is, never justifies known concepts, and every sentence adds non-obvious policy ("Do not run that command autonomously", "Do not probe first with --help, --version, devices, appstate, snapshot, or screenshot"). Matches the level-5 anchor (lean, assumes Claude's competence, every token earns its place); level 4 would require some trimmable over-explanation, and none is present. | 5 / 5 |
Actionability | Concrete, executable commands are given for the main path — `npm install -g agent-device@latest` and the copy-paste-ready `agent-device open <app-or-bundle-id> --platform ios --foreground` — plus five exact help subcommands. However, the core interaction verbs advertised in the description (tap, type, scroll, snapshot reading) never get a syntax example in the body; their shapes are delegated to the CLI's snapshot hints and help topics. That is a deliberate, defensible design, but it leaves a minor gap versus the level-5 anchor ('specific examples cover the common cases'), so level 4 ('mostly executable, minor gaps') is the best fit. | 4 / 5 |
Workflow Clarity | The sequence is clear and conditional: start immediately for normal tasks → target iOS explicitly on open → "Follow the initial interactive snapshot and corrective error hints" → escalate to a version-matched help topic only when specialized. There is an explicit error-recovery path ("If the shell reports that agent-device is unavailable, stop and ask the user") and a built-in feedback loop (snapshot → act on corrective hints). No destructive/batch operation exists, so the validation cap does not apply, and the simple-skill exception holds: the primary action is unambiguous. Anti-drift check against level 4 ('minor validation gaps'): validation is explicitly delegated to the tool's corrective hints and stated as such, so 5 is the better fit. | 5 / 5 |
Progressive Disclosure | The skill is under 50 lines with no bundle files (references/, scripts/, assets/ do not exist), and nothing in the body belongs in a separate file. Detail is one level deep via `agent-device help <topic>` — clearly signaled and gated ("Only when the task is specialized or a command shape is unclear") — which is exactly the well-signaled, one-level-deep reference pattern of the level-5 anchor. Level 4 would require organization gaps, and the flat 29-line body with its condition-gated escalation reads cleanly without headers. | 5 / 5 |
Total | 19 / 20 Passed |