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 core Workflow, Output Format, and Requirements sections are concrete and actionable, and detail is appropriately offloaded to a real one-level-deep reference file. However, three boilerplate sections ('When to Use', 'Key Features', 'Implementation Details') pad the body with generic filler, the workflow lacks embedded validation checkpoints, and two file/section pointers are broken or misspelled.
Suggestions
Delete or drastically trim the 'When to Use', 'Key Features', and 'Implementation Details' boilerplate sections that restate the description and add generic filler like 'Execution model: validate the request...'; keep the body focused on the Workflow, Output Format, and Requirements content.
Add an explicit validation step to the Workflow (e.g., check events <= sample size per group, confirm all intervention groups and outcomes were captured, and flag blank values for user review) instead of relying on the implicit 'Review the generated output' in the run plan.
Fix broken navigation: correct 'extraction-promots.md' to 'references/extraction-prompts.md', correct 'See `## Workflow` above' (the section is below), and replace the non-portable `cd "20260316/scientific-skills/..."` example path with one relative to the skill directory; also name the actual PMID database used for lookup.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Several padded, boilerplate sections: 'When to Use' restates the frontmatter description nearly verbatim plus generic filler ('when a data analytics task needs a packaged method instead of ad-hoc freeform output', 'when the user expects a concrete deliverable'), 'Key Features' repeats the description and vague claims, and 'Implementation Details' offers generic lines like 'Execution model: validate the request, choose the packaged workflow, and produce a bounded deliverable'. Not anchor 1 because the Workflow/Output Format/Requirements sections are tight and non-redundant, but the filler is too extensive for anchor 3's 'some unnecessary explanation'. | 2 / 5 |
Actionability | Concrete, executable guidance: a runnable script with usage ('python scripts/extract_pdf.py --help'), a fully specified JSON output schema, and exact per-type field lists (n/event; n/mean/sd; n/HR/95% CIs). Minor gaps keep it below anchor 5: the 'PMID database' is never identified, the example `cd "20260316/scientific-skills/..."` path is not a real usable path, and the pointer to the detailed prompts misspells the filename ('extraction-promots.md' vs the actual 'references/extraction-prompts.md'). | 4 / 5 |
Workflow Clarity | The Workflow section has a clear 5-step sequence with an explicit branch (PMID provided -> database lookup first; otherwise LLM extraction), but checkpoints are missing or implicit: no validation step inside the workflow itself (e.g., verifying events <= n or cross-checking totals against reported sample sizes), and verification appears only as an implicit 'Review the generated output' in the run plan. Fits anchor 3 rather than 4's 'most checkpoints present'. | 3 / 5 |
Progressive Disclosure | Good structure: detailed per-outcome-type prompt/JSON structures are correctly deferred to the one-level-deep 'references/extraction-prompts.md' (verified to exist), while SKILL.md keeps a concrete overview schema. Minor gaps: the closing pointer misspells the filename ('extraction-promots.md'), the loud 'See references' note sits at the very bottom, and 'See `## Workflow` above' is a broken internal pointer (that section appears below, not above). | 4 / 5 |
Total | 13 / 20 Passed |