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 exceptionally actionable, with concrete commands, verified tool behavior, explicit validation checkpoints and error-recovery loops, but it pays for that with poor token economy: concept explanations Claude already knows, triple-stated conventions, and ~400 lines of tool tables duplicated from the reference files. The reference bundle itself is well-structured and one level deep, but the SKILL.md has not actually moved the bulk detail out to it, and one cited file (QUICK_START.md) is missing.
Suggestions
Move the ToolUniverse tool tables (EVA, Variant Notation Conversion, Ensembl ID Utilities/Liftover/LD, Phenotype) into the existing reference files and keep only a one-row-per-tool 'when to use' summary with pointers in SKILL.md — the references already claim to hold the full tables, so the inline duplication is pure token cost.
State the coding-only-denominator rule once (e.g., in the PRIMARY SCRIPTS/variant_fraction.py row) and reference it elsewhere instead of repeating it in the CRITICAL section and again in Analysis Conventions; likewise drop background Claude already knows (LD definition, SpCas9 PAM orientation, VCF quality-filtering rationale).
Fix the dangling 'QUICK_START.md' reference — either add the file to the bundle or remove the pointer — since a cited-but-absent resource breaks navigation at exactly the point a reader needs it.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~800-line body includes several sections of knowledge Claude already has — "LD measures non-random co-inheritance of alleles at different loci", CRISPR "PAM sequence (NGG for SpCas9) must lie 3' of the target", a generic VCF quality-filtering rationale, and definitions of SPDI/HGVS/VCF notation — and it is noticeably redundant: the coding-only-denominator rule appears three times (PRIMARY SCRIPTS table, the 'CRITICAL' section, and Analysis Conventions), and the notation-conversion parameter tables are inlined even though references/variant_notation_conversion.md already holds the 'full parameter tables'. This matches level 2 ('noticeably verbose; several unnecessary explanations or padded sections'); it is not level 1 because a large share of the content is genuinely valuable, non-obvious, live-verified tool knowledge. | 2 / 5 |
Actionability | Copy-paste-ready bash invocations with concrete flags ("--workdir /tmp/hc_run --sample-name SAMPLE", "--pattern '*.xlsx' --header-rows 2"), python snippets with expected return values, named output keys to grep (SNP_COUNT_ALLELES, AVERAGE_PER_SAMPLE), tool call signatures with parameters and response shapes, and decision guides ('start from what you have, land on what you need'). This fully meets the level-5 anchor for executable, common-case-covering guidance. | 5 / 5 |
Workflow Clarity | A clear ordered decision sequence exists (RULE ZERO pre-computed-results check → PRIMARY SCRIPTS → phase recipe), phases 1-7 are explicitly sequenced, and there are explicit validation checkpoints and feedback loops: scripts emit 'PLOIDY=<value> from the VCF header so you can confirm', a sanity check ('If your "synonymous fraction" is much lower than ~0.4 ... restrict to CODING and recompute'), retry-with-longer-timeout guidance for flaky endpoints, and the mandatory 'ALWAYS report all THREE proportions' checklist. Batch/destructive-adjacent operations (annotation batches, pipeline writes) do have verification steps, so no cap applies. | 5 / 5 |
Progressive Disclosure | References are real, one level deep, and clearly signaled ('See references/vcf_filtering.md for all filter options'), and all cited scripts exist. However the body inlines hundreds of lines of ToolUniverse tool reference tables (EVA, notation conversion, Ensembl utilities, phenotype) that the skill itself says are fully documented in the reference files — content that clearly belongs in separate files is inline, matching level 3 ('content that should be separate is inline'). Additionally, 'QUICK_START.md' is cited under Additional Resources but is absent from the bundle, a dangling reference. | 3 / 5 |
Total | 15 / 20 Passed |