Content
63%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 dense, actionable workflow with exact tool calls and quantitative scoring, weakened by broken bundle references and some re-taught background biology. The core defect is that four linked files the workflow repeatedly depends on do not exist in the skill directory.
Suggestions
Ship the four referenced bundle files (SCORING_TABLES.md, INPUT_REFERENCE.md, REPORT_TEMPLATE.md, TOOLS_REFERENCE.md) or inline their essential content — the cancer-specific thresholds, ORR baselines, and verified tool parameters the workflow depends on are currently unreachable.
Trim the 'Reasoning Before Searching' section to the scoring-relevant caveats (e.g., RCC responding despite low TMB) and remove the duplicate resistance-gene listing and the trailing URL list to save tokens.
Add parameters for tools currently called without them (iedb_search_epitopes, civic_search_evidence_items, and the Phase 6 gene-expression queries) so every tool invocation is fully specified.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The 'Reasoning Before Searching' section re-teaches well-known oncology biology ('Caused by defective DNA mismatch repair (MMR)', 'The direct target of pembrolizumab'), resistance genes are listed twice (KEY PRINCIPLES #6 and Phase 7), and the trailing database URL list adds little — the document could be tightened despite being mostly dense and actionable. | 3 / 5 |
Actionability | Highly concrete guidance: exact tool calls with parameters (e.g., "fda_pharmacogenomic_biomarkers(drug_name='pembrolizumab')", "MyGene_query_genes: use query (NOT q)"), explicit point values (STK11 -10, POLE +10), and a fully specified score formula with floor and cap. Minor gaps: several tools (iedb_search_epitopes, civic_search_evidence_items, Phase 6 expression queries) are invoked without parameters, and the promised parameter reference file does not exist. | 4 / 5 |
Workflow Clarity | Eleven clearly sequenced phases with explicit checkpoints ('LOOK UP DON'T GUESS', 'BEFORE calling ANY tool, verify parameters', flag unknowns instead of defaulting to moderate) and confidence tiers tied to biomarker completeness. No error-recovery or feedback loops for failed or empty tool calls, keeping it below a 5; no destructive or batch operations apply the cap. | 4 / 5 |
Progressive Disclosure | The body itself is well organized with one-level, clearly signaled references, but all four referenced files (INPUT_REFERENCE.md, SCORING_TABLES.md, REPORT_TEMPLATE.md, TOOLS_REFERENCE.md) are missing from the bundle, so the cancer-specific thresholds, ORR baselines, report template, and tool-parameter tables the workflow depends on are unreachable. | 3 / 5 |
Total | 14 / 20 Passed |