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.
A well-engineered operational skill: lean, cancer-specific, and packed with concrete tool calls, point-valued scoring, and an explicit phase sequence with parameter-verification checkpoints. Its one material defect is that every referenced bundle file is missing, so the deferred detail the workflow depends on (scoring tables, tool parameter reference, report template) cannot actually be loaded.
Suggestions
progressive_disclosure: Ship the four referenced bundle files (INPUT_REFERENCE.md, SCORING_TABLES.md, REPORT_TEMPLATE.md, TOOLS_REFERENCE.md) in the skill directory — every phase links to them, and the scoring tables and tool parameter tables they should contain are load-bearing for Phases 1, 2, 5, 8, and 11.
progressive_disclosure: If the bundle files cannot be provided, inline the minimal essentials (cancer-type ORR baselines, PD-L1 TPS/CPS cutoffs, and the ICI drug profile table) into SKILL.md and drop the dangling links.
workflow_clarity: Add explicit error-recovery feedback loops for tool failures — e.g., what to do when a gene ID cannot be resolved via MyGene_query_genes or when FDA/civic lookups return no evidence — so the 11-phase sequence degrades gracefully instead of implicitly assuming every call succeeds.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and operational — point-valued rules ('STK11 (-10), PTEN (-5), JAK1/2 (-10 each), B2M (-15)'), parameter gotchas ('use `query` (NOT `q`)'), and thresholds — with little generic explanation. Minor over-explanation remains: the 'Reasoning Before Searching' section re-states neoantigen biology Claude already knows, and the 9-item 'KEY PRINCIPLES' list repeats rules stated later in the phases. Efficient with minor trimming possible, matching anchor 4 rather than the fully lean anchor 5. | 4 / 5 |
Actionability | Mostly executable guidance: concrete tool calls with parameters ('fda_pharmacogenomic_biomarkers(drug_name=\'pembrolizumab\')', 'HPA_get_cancer_prognostics_by_gene(gene_name=\'CD274\')'), an explicit scoring formula ('missense_count * 0.3 + frameshift_count * 1.5'), and a fully specified Phase 10 score integration. Minor gaps keep it below anchor 5: several tool calls lack full parameter signatures (deferred to TOOLS_REFERENCE.md) and there is no runnable Python snippet for the score computation despite the 'COMPUTE, DON'T DESCRIBE' mandate. | 4 / 5 |
Workflow Clarity | The 11-phase sequence is clearly ordered with explicit input requirements, a deterministic scoring integration, and checkpoints such as 'BEFORE calling ANY tool, verify parameters', 'LOOK UP DON'T GUESS', and 'Do not default to "moderate" for unknowns — flag them explicitly'. Below anchor 5 because there are no explicit error-recovery feedback loops (e.g., what to do when a tool call fails or a gene ID cannot be resolved); above anchor 3 because checkpoints are explicit rather than implicit. | 4 / 5 |
Progressive Disclosure | The structure pattern is good — overview phases each pointing to one-level-deep references at the point of need ('See [SCORING_TABLES.md](SCORING_TABLES.md)', '[INPUT_REFERENCE.md](INPUT_REFERENCE.md)', '[TOOLS_REFERENCE.md](TOOLS_REFERENCE.md)', '[REPORT_TEMPLATE.md](REPORT_TEMPLATE.md)') — but none of these four referenced files exist in the bundle (no references/, scripts/, or assets/ directories are present). Navigation is clearly signaled yet broken at runtime, and the deferred content (scoring tables, tool parameters) is inaccessible, which is a larger gap than anchor 4's 'minor organization gaps'. | 3 / 5 |
Total | 15 / 20 Passed |