Content
53%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.
The body is well-structured with good routing tables, an explicit output template, and honest anti-patterns, but its executable core is thin: the pivotal Perplexity query exists only as commented pseudocode, no step validates API success or page writes, and several sections (the conformance-test stubs, the empty cron example) consume tokens without adding guidance. It reads as a solid design doc one pass away from an operational skill.
Suggestions
Make the Invocation example executable: show one complete, filled-in Perplexity request (real curl with an actual composed prompt containing brain context and the 'find what's NEW since YYYY-MM-DD, cite every claim' instruction) instead of the current '# 2. ...' comment sketch.
Add validation checkpoints to the workflow: check the API response for citations before writing, confirm `gbrain put research/<slug>` succeeded, and define a retry/flag-for-review path when Perplexity returns nothing new or contradicts brain knowledge (especially for the cron-driven deal monitoring pattern).
Trim token-inefficient sections: replace the empty 'Deal / company monitoring' comment-only bash block and the two conformance-test stub sections ('Contract', 'Output Format') with a single line pointing to the conformance contract, and drop the 'key insight' paragraph in favor of one sentence.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly efficient sections ("What this does", "Models", "Anti-Patterns"), but several sections pad without guiding: "The key insight: Perplexity doesn't just search — it reads and synthesizes with citations" over-explains rationale Claude can infer, the "Deal / company monitoring" bash block contains only a comment ("# Weekly: pull recent news per company; flag changes for review"), and the trailing "Contract"/"Output Format" sections explicitly exist "for the conformance test" rather than the reader. This is noticeably below anchor 4's 'minor instances that could be trimmed' and above anchor 2's pervasive padding. | 3 / 5 |
Actionability | The central Invocation section is comment-pseudocode ("# 2. Compose the Perplexity query with brain context inline" with a triple-quoted sketch, and a curl example whose message content is literally "...") rather than executable code, matching anchor 3's 'pseudocode instead of executable code; missing key details'. Real commands exist (gbrain get/query/put, the curl skeleton) and the output template is concrete, so it is clearly above anchor 2's high-level-hints-only, but the actual query the skill hinges on is never shown — below anchor 4's 'mostly executable'. | 3 / 5 |
Workflow Clarity | The Invocation section lists a clear 5-step sequence (pull context → compose query → call API → put_page → cross-link), but there are no validation checkpoints: no check that the Perplexity call succeeded or returned citations, no verification the research page was written, and no error-recovery loop for a mutating, batch-oriented skill (mutating: true, cron-driven deal monitoring). This lands exactly on anchor 3 ('sequence present but checkpoints missing') — above anchor 2's rough sequence with undefined steps, below anchor 4's most-checkpoints-present. | 3 / 5 |
Progressive Disclosure | Sections are well-organized with clear headers, a routing table, and one-level-deep, clearly signaled references ("see [conventions/quality.md](../conventions/quality.md)", "see [conventions/brain-first.md](../conventions/brain-first.md)"); no bundle files exist, so everything navigable is signaled inline. Minor gap: the ~30-line output page template and the integration patterns are inlined where a reference file would keep SKILL.md leaner — 'good structure; most content appropriately placed; minor organization gaps' rather than anchor 5's ideal split. | 4 / 5 |
Total | 13 / 20 Passed |