Content
71%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 unusually dense, high-discipline runner skill: concrete commands, explicit validation checkpoints, honest fallback semantics, and no filler. Its weakness is structural — roughly a hundred lines of recording and intent-spec rules are inlined in SKILL.md rather than split into references, and the intent-spec prose is compressed to the edge of ambiguity.
Suggestions
Move the Record video and Intent specs sections into one-level-deep reference files (e.g. references/record-video.md, references/intent-specs.md), keeping a short summary plus the shared rules pointer inline, so the runner body stays an overview.
Rewrite the Intent-specs replay/heal decision as a small ordered list or table instead of nested prose conditionals, mirroring the clarity of the first-match grading table it defers to.
Fix the tool-load mismatch: add `javascript_tool` to the ToolSearch select list (or state explicitly how it is loaded), since the locator ladder, semantic assertions, and captions all depend on it.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense operational instruction with almost no explanation of known concepts — every sentence issues commands like "Call `tabs_context_mcp`. If it errors or returns no browser, the extension is not connected. **Do not improvise.**". It falls short of 5 because of repetition (the playwright fallback is restated in Preflight, Auth, and Hard Rules) and a few justification flourishes ("That see→act loop is the point: no throwaway batch script, no blind compile") that could be trimmed. | 4 / 5 |
Actionability | Mostly copy-paste ready: the full `ToolSearch "select:..."` string, exact YAML fallback verdict blocks, concrete `node $CAP start/step/stop` commands, and an input-resolution table. Not 5 because of small gaps — e.g. `javascript_tool` is used in the locator ladder and captions but is absent from the ToolSearch load list, and some prose ("re-judge the failed assertions on the current page when every route step ran") leaves the exact command to inference. | 4 / 5 |
Workflow Clarity | The sequence is explicit with real checkpoints: numbered Critical First Actions, a preflight that validates the extension connection, per-action `read_page` re-reads, auth detection with an honest-inconclusive rule, bail-mode semantics, and feedback loops (locator healing, exit-6 gif_creator fallback chain, lessons loop). Not 5 because the Intent specs section compresses its replay/heal/grade rules into nested prose conditionals that a reader must unwind, and grading defers to a contract table not shown here. | 4 / 5 |
Progressive Disclosure | References out to the contract and sibling files are well signaled ("**Read the spec-run contract first.** It owns the locator ladder..."), but no bundle files exist and the file is a ~320-line monolith: the Record-video (~40 lines) and Intent-specs (~60 lines) sections are self-contained rule sets that clearly belong in one-level-deep reference files. Between anchor 3 (content that should be separate is inline) and 4; the inlined bulk keeps it at 3. | 3 / 5 |
Total | 15 / 20 Passed |