Content
78%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 disciplined router skill: it stays tiny, gates setup with an explicit version check and failure path, and deliberately delegates volatile details to `agent-device help dogfood`. The gaps are the pseudocode-style loop line, an inline time-sensitive version floor, and underspecified evidence-capture steps in the body itself.
Suggestions
Expand the arrow-notation loop ("open app -> snapshot -i + screenshot -> ...") into a numbered step list with one concrete example command per step (e.g., the actual open/snapshot invocations), keeping `help dogfood` as the source of full detail.
Move the "agent-device >= 0.14.0" floor into a short, clearly-labeled compatibility/version-note section (or fold it into the `help dogfood` guidance step) so time-sensitive details are quarantined from the evergreen workflow.
State the minimal in-body evidence expectations for an issue (e.g., what a finding record must contain) so report shape does not depend entirely on the deferred CLI help.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean with no concept explanations Claude already knows — every line is setup gating, commands, or workflow. However, the time-sensitive version floor "Require `agent-device >= 0.14.0`" sits inline rather than in an old-patterns/deprecated section, which the guidelines say should penalize conciseness; otherwise this would be a 5. | 4 / 5 |
Actionability | Quotes: "agent-device --version" and "agent-device help dogfood" — two executable commands with explicit failure handling. But the core workflow is arrow notation ("open app -> snapshot -i + screenshot -> explore flows -> capture evidence per issue -> close") rather than copy-paste commands, so it is 'mostly executable with minor gaps' (level 4), not fully copy-paste ready (level 5). The deferral of exact commands to `help dogfood` is explicitly justified, which keeps this above level 3. | 4 / 5 |
Workflow Clarity | Quotes: "If that fails, stop and tell the user to expose a trusted `agent-device` binary" and "If older, stop and tell the user to upgrade" — a clear setup→guidance→loop sequence with an explicit validation checkpoint and failure path. It stays at level 4 rather than 5 because the in-loop steps ("explore flows", evidence capture, report shape) are delegated without inline checkpoints; the destructive/batch validation cap does not apply since this is read-only exploration. | 4 / 5 |
Progressive Disclosure | The body is ~27 lines with no bundle files, well-organized into router intro, version gate, guidance, and loop sections; per the simple-skill exception, an under-50-line skill with no need for external references can score 5 with well-organized sections. Its single external reference ("agent-device help dogfood") is one level deep and clearly signaled. | 5 / 5 |
Total | 17 / 20 Passed |