Conciseness

The ~500-line body has substantial duplication and background Claude already knows: the R-vs-pydeseq2 guidance appears in four places ("Use R DESeq2, not pydeseq2", "Analysis conventions", "Known Limitations", "Dispersion estimates: R DESeq2 vs pydeseq2 diverge"), the union-denominator rule appears three times, "Also DE in strain X" and "uniquely DE" rules each appear twice, and sections like "Core Principles" ("Data-first", "Statistical rigor") and the "Interpretation Framework" table (padj<0.05 = significant, LFC>1 = 2x) restate textbook knowledge. This is noticeably below the mostly-efficient anchor-3; not anchor 1 because the majority of the material (script contracts, notebook-filter pitfalls) is genuinely non-obvious and earns its tokens.

Actionability

Highly concrete: full CLI invocations with arguments for all six primary scripts (e.g., the r_deseq2_wrapper.py multi-contrast example), parseable output-line formats, R code for dispersion queries, tu run commands, and decision tables mapping question phrasing to output lines. Falls short of 5 because the executable guidance is internally conflicted — the "Analysis Workflow" Step 3 directs to "Run PyDESeq2" and pydeseq2_workflow.md while the CRITICAL section commands "Use R DESeq2, not pydeseq2" — leaving the reader unsure which executable path is sanctioned.

Workflow Clarity

A 7-step workflow exists (parse question → load/validate → inspect metadata → run → filter → dispersion → enrichment), each step linked to a reference file, with an Error Quick Reference table for recovery. But checkpoints are implicit rather than explicit, the CRITICAL "read the executed notebook FIRST" precedence rule and the primary-scripts track are not integrated into that workflow (they form a second, conflicting track), the CRITICAL numbered list has two items numbered "3", and the pydeseq2-vs-R contradiction breaks the single clear sequence — matching anchor 3 rather than 4.

Progressive Disclosure

References are one level deep and each link carries a one-line description, and all 9 scripts exist — but 2 of the 12 referenced files are missing from references/ (design_formula_guide.md, referenced at the metadata step and in the file list; r_clusterprofiler_guide.md), so those links are broken. Additionally, hundreds of lines of edge-case heuristics (the entire "LOOK UP DON'T GUESS" and "Analysis conventions" sections) are inlined in SKILL.md rather than split into reference files, which is the anchor-3 pattern of content that should be separate living inline.

Specificity

Lists multiple concrete actions — "DEG lists, fold changes, dispersion estimation, design formulas including covariates, multi-condition contrasts, and Venn-set operations" — plus explicit routing rules ("edgeR (QL-F / exact test for small replicate counts)", "limma-voom (large n / complex designs)"), which is anchor-5 coverage. However, the trigger clause "Use when you have a count matrix + metadata" is second-person voice, and the rubric requires reducing the specificity score by 1 for second-person phrasing, bringing it to 4.

Completeness

Explicitly answers both: what — "RNA-seq differential expression analysis with DESeq2, edgeR, and limma-voom — DEG lists, fold changes, dispersion estimation, design formulas... multi-condition contrasts, and Venn-set operations"; and when — "Use when you have a count matrix + metadata, want to find DEGs, or need dispersion/PCA/clustering analysis". This matches the anchor-5 example structure (concrete what + explicit when with trigger phrases), and is above anchor 4 because the when-clause lists three concrete triggering situations rather than a generic one.

Trigger Term Quality

Natural user terms are well covered: "RNA-seq", "differential expression", "DESeq2", "edgeR", "limma-voom", "DEGs", "count matrix", "dispersion", "PCA", "clustering". Not anchor 5 because common synonyms and trigger variants are missing — e.g., "differentially expressed genes" spelled out, "enrichment" (a capability covered in the body), "batch effect", and file extensions like .csv/.txt — so a few natural phrasings a user might say would not lexically match.

Distinctiveness Conflict Risk

Clear niche — differential expression on RNA-seq count matrices with three named frameworks and dataset-specific triggers ("count matrix + metadata", "DEGs", "executed.ipynb") — so it would not fire for general statistics, general bioinformatics, or non-RNA-seq tasks. Distinct, domain-specific tool names give it minimal overlap risk with other skills.