Content
70%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 highly actionable for a tool-orchestration skill — exact tool names, parameter gotchas, conditional phases, and fallback loops — and the workflow is unambiguous. Its weaknesses are repetition and inlined domain tutorials that pad the token budget, and the absence of any progressive disclosure: four example workflows and detailed phase content all live inline in a single long SKILL.md with no reference files.
Suggestions
Consolidate the ALPHA_GENOME_API_KEY fallback, which is currently explained in full in Phase 0.5, Phase 4.5, and Fallback Strategies — state it once in Phase 0.5 and reference it elsewhere.
Cut domain knowledge Claude already has (histone mark meanings in the 'Is the variant in a regulatory element?' section, eQTL tissue-specificity interpretation in Phase 2) down to the decision-relevant minimum.
Move the four example workflows into a references/EXAMPLES.md and link to it from a short 'Example Workflows' pointer, keeping SKILL.md as a lean overview of the phases.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Quotes: "H3K27ac signals active enhancers and active promoters; H3K4me1 alone marks poised enhancers; H3K4me3 marks active promoters; H3K27me3 marks silenced regions" and "A tissue-specific eQTL suggests cell-type-specific regulation; a ubiquitous eQTL suggests a core regulatory element" — didactic domain knowledge Claude already has. The ALPHA_GENOME_API_KEY fallback is explained three times (Phase 0.5: "If it isn't configured, the AlphaGenome tools won't appear...", Phase 4.5: "If it's unavailable (no key)...", Fallback Strategies: "AlphaGenome tools aren't in your toolset: ..."), and the ASCII workflow diagram restates the phase headers. Not 4: the padding is more than minor; not 2: the bulk is dense, non-obvious tool guidance (param gotchas, API quirks) that earns its tokens. | 3 / 5 |
Actionability | Quotes: "`EnsemblVEP_annotate_rsid` (param is `variant_id`, not `rsid`)", "Use `assay_title="TF ChIP-seq"` (not just "ChIP-seq")", "Use `p_value=5e-8` for genome-wide significance", and concrete example calls like "RegulomeDB_query_variant(rsid="rs429358")". Mostly executable, tool-accurate guidance with exact parameter names and values. Not 5: example workflows 3 and 4 use placeholder arguments ("chromosome=..., position=...") and no end-to-end runnable form is shown; minor gaps remain. | 4 / 5 |
Workflow Clarity | Quotes: an explicit phase sequence (Phase 0 through Phase 6) with conditional branches — "Run this phase when either condition holds: (a) Phases 1-4 came back empty or weak... or (b) Phase 0.5's AVI_SCORE was high" — plus error-recovery feedback loops in Fallback Strategies ("GWAS Catalog returns empty: Switch from free-text `disease_trait` to `efo_id`") and validation checkpoints ("Use `GTEx_get_median_gene_expression` to confirm that the target gene is actually expressed... before placing weight on eQTL evidence"). Matches the score-5 anchor: clear sequence, explicit checkpoints, feedback loops for error recovery. Not a destructive/batch skill, so no cap applies. | 5 / 5 |
Progressive Disclosure | No bundle files exist (no references/, scripts/, or assets/ directories) and everything is inlined in a ~280-line SKILL.md, including four full example workflows ("### GWAS Variant Functional Annotation (rs429358 / APOE)", "### Annotation-Silent Variant Resolved via Sequence Prediction", etc.) that would fit naturally in a separate EXAMPLES.md. Section headers are clear, so structure exists, but content that should be split out is inline with no one-level-deep references — matching the score-3 anchor ('content that should be separate is inline'). Not 4: there is no reference structure at all to be 'mostly clear'; not 2: the document is well-sectioned and easy to navigate, not a wall of text or buried references. | 3 / 5 |
Total | 15 / 20 Passed |