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.
A highly actionable, well-sequenced CLI guide with real error-recovery loops and a sound reference structure. Its main weakness is length: duplicated WebMCP prose and inlined reference-grade detail (dashboard, read flags) bloat the overview, and one cited bundle path (templates/*) does not exist.
Suggestions
Deduplicate the WebMCP paragraphs: keep the discovery/security details in one place (e.g. a references/webmcp.md or the core-loop section) and have the Troubleshooting entry reference it instead of repeating ~80% of the same text.
Move reference-grade configuration detail — dashboard allowed-origins/token behavior, the read command's flag-by-flag semantics, and idle-timeout/restore-save policy specifics — into references/*.md files, leaving SKILL.md as a lean overview with one example each.
Fix the dangling 'templates/*' citation in the 'Full reference' section: either ship the starter shell scripts in the bundle or remove the bullet until they exist.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body contains no explanations of concepts Claude already knows — every sentence is agent-browser-specific — but at ~580 lines it is noticeably longer than an overview needs to be. The WebMCP announcement/security paragraph appears nearly verbatim in both 'The core loop' and 'Troubleshooting', and the 'read' flag enumeration plus the dashboard origin/token paragraphs are dense reference-grade prose that could be tightened or moved to references. This fits 'mostly efficient but could be tightened' rather than the noticeably-padded 2 anchor. | 3 / 5 |
Actionability | Nearly every section is copy-paste-ready bash: a working quickstart, a login workflow with concrete refs, semantic-locator and CSS fallbacks, a wait-strategy menu with exact commands, heredoc-based eval examples, and doctor diagnosis commands. Specific examples cover the common cases, matching the fully-executable 5 anchor. | 5 / 5 |
Workflow Clarity | The core loop is explicitly sequenced (open → snapshot -i → act on refs → re-snapshot), the waiting section prescribes a concrete post-action checkpoint ('After any page-changing action, pick one'), and the troubleshooting section is a set of error-recovery feedback loops (e.g. 'Ref not found' → re-snapshot → use new refs). Destructive repairs are gated behind 'doctor --fix', so the destructive-operation cap does not apply. | 5 / 5 |
Progressive Disclosure | References are one level deep, clearly signaled, and all ten cited references/*.md files exist in the bundle. However, reference-grade material is inlined in SKILL.md (dashboard origin/token configuration, WebMCP discovery details), and 'templates/*' is cited under 'Full reference' with no templates directory present in the bundle — minor organization gaps that keep this below the 5 anchor. | 4 / 5 |
Total | 17 / 20 Passed |