Content
81%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 exceptionally actionable, well-sequenced operational guide with real validation feedback loops and a properly structured bundle. Its weakness is length: duplicated warnings and rhetorical troubleshooting narrative inflate the body, and some of that detail belongs in the reference files rather than the main SKILL.md.
Suggestions
State the never-relaunch-the-browser rule once — keep the full rationale in either Step 0 or the Gotchas section and cross-reference it from the other, and trim the rhetorical framing around CDP unavailability to the operational facts (probe every time, never cache, fall back on exit 3).
Move the AppleScript routing trap narrative (the ~20-line section including the observed windows=1 tabs=1 anecdote) into references/applescript_fallback_channel.md, leaving in SKILL.md only the decision rule: check probe's chrome_instances.automation list before falling back to AppleScript.
Collapse the duplicated channel-specific download examples into one parameterized example plus a short channel-substitution note, keeping the CDP worked example and pointing AppleScript users to the channel reference.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with non-obvious empirical detail Claude does not know (exit-code contracts, endpoint shapes, Chrome 136 CDP hardening), but several passages could be tightened: the never-relaunch-the-browser warning is stated in full twice (Step 0 and Gotchas), CDP unavailability gets three paragraphs of rhetorical framing ("Assume you are being lied to by default"), and the AppleScript routing trap runs ~20 lines of narrative that belongs in its reference file. This sits at 'mostly efficient but includes some unnecessary explanation or could be tightened', above level 2 since nothing explains concepts Claude already knows. | 3 / 5 |
Actionability | Every step carries copy-paste-ready commands: the exact ToolSearch select string, `uv run --with websockets python scripts/cdp_channel.py probe/eval`, full endpoint URLs with query parameters (`rendering_mode=messages&render_all_tools=true`), worked download-and-base64-decode loops for two channels, and concrete verification (`file <name>` magic bytes, byte-size comparison against `files[].size_bytes`). Common cases are covered end to end. | 5 / 5 |
Workflow Clarity | Steps 0–5 are clearly sequenced with decision tables (channel order, account cases A/B/C) and an exit-code contract for fallback. Validation is explicit and layered: the fidelity gate refuses to write lossy output and names the offending block, the regression suite must run after renderer changes, and the completion check reads the `files:` stderr line so file downloads are not skipped. Failures feed back into fixes (teach the renderer the new shape, re-run selftest). | 5 / 5 |
Progressive Disclosure | References are one level deep, real files (both exist on disk), and clearly signaled at point of need; scripts are properly separated and linked. However, the ~450-line SKILL.md inlines substantial troubleshooting narrative — the AppleScript routing trap, the multi-paragraph CDP availability discussion, and duplicate channel-specific download examples — that belongs in the two reference files, so content that should be split is inline (anchor 4, not 5). | 4 / 5 |
Total | 17 / 20 Passed |