Content
67%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 delivers a concrete, well-sequenced verification workflow with an executable command, a complete output template, specific error-recovery loops, and clearly signaled one-level-deep references. Its main weakness is redundancy — verdict definitions, purpose, and trigger contexts are each stated twice, plus two conformance-test boilerplate sections — which costs tokens without adding guidance.
Suggestions
Merge the 'Standards (the rigor bar)' verdict definitions into Step 4's verdict list (or vice versa) so each of the five verdicts is defined exactly once, keeping the stricter evidence criteria from Standards.
Delete or compress the 'Contract' and 'Output Format' sections, whose own text admits they exist 'for the conformance test' — replace with a single line pointing to the conformance suite, freeing roughly 15 lines.
Move the brain-page output template and the databases table into a references/ file (e.g. references/output-template.md) and reference it one level deep, trimming SKILL.md to a lean overview.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The five verdict definitions appear twice — once in Step 4 ('Verified — claim is accurate; raw data available; replication exists...') and again nearly verbatim in 'Standards (the rigor bar)' — and 'What this is'/'When to use this' restate the frontmatter description. The 'Contract' and 'Output Format' sections are explicitly conformance-test boilerplate ('this section exists for the conformance test') that adds tokens without execution value. This matches 'Mostly efficient but includes some unnecessary explanation or could be tightened'; not 2 because most sections (workflow, template, failure handling, database table) are substantive and non-padded. | 3 / 5 |
Actionability | Concrete, executable elements abound: 'gbrain query "<paper title> OR <author name> OR <claim keywords>"', a full copy-paste brain-page template with frontmatter, URL patterns per database ('retractionwatch.com/?s=NAME'), and specific error semantics ('poll gbrain write-request <request_id> instead of writing a second verdict page'). Matches 'Mostly executable guidance; concrete code or commands with minor gaps'. Not 5 because the perplexity-research prompt is specified as a requirements list rather than a verbatim template, and the claim-slug naming convention is left implicit. | 4 / 5 |
Workflow Clarity | The five steps (Scope → Brain-first lookup → perplexity-research → Format verdict → Cross-link) are clearly sequenced, and 'When it fails' provides explicit feedback loops ('on rate_limited retry once after the stated delay', 'retry with the author, journal or year before saying the brain has no prior notes'). The Standards section acts as a per-verdict evidence bar. Matches 'Clear sequence with most checkpoints present; minor validation gaps'. Not 5 because there is no explicit check that the perplexity response actually contains all five required trace elements before the verdict is formatted and written. | 4 / 5 |
Progressive Disclosure | No bundle files exist (no references/, scripts/, or assets/), and all file references are one level deep and clearly signaled at point of use ('see [conventions/quality.md](../conventions/quality.md) for citation rules', '[perplexity-research](../perplexity-research/SKILL.md)'). Sections are well-organized with a scannable structure. Matches 'Good structure; most content is appropriately placed; references mostly clear; minor organization gaps'. Not 5 because the ~45-line output template and the databases table are inlined in SKILL.md where a references/ file could keep the overview leaner. | 4 / 5 |
Total | 15 / 20 Passed |