Conciseness

The body is dense and almost every token is version-specific, non-obvious knowledge (the vanilla-vs-change DE column trap, the ArrowStringArray write failure, the removed use_gpu kwarg) rather than concepts Claude already knows. It is not a 5 because the Gotchas and Troubleshooting tables repeat the same three errors (use_gpu TypeError, IORegistryError, KeyError 'lfc_mean') and the "This is a pure skill ... There is no host runtime and no LLM API" paragraph is environment meta-commentary that could be trimmed.

Actionability

The scVI, scANVI, and DE blocks are fully executable with exact kwargs and the exact output-column list, and the troubleshooting table maps concrete error strings to fixes. It is not a 5 because the Modal pipeline.py snippet is a placeholder ("# ... setup_anndata / scVI / scANVI / DE — see the recipe above ...", "paste the helper into THIS script (below)") and its h5ad_safe_obs call depends on a helper whose definition is nowhere in the skill — the referenced kernel.py does not exist in the bundle.

Workflow Clarity

The pipeline is clearly sequenced with inline ordering guards ("preserve raw BEFORE any normalize/log1p", stash counts layer, then setup_anndata → train → latent → neighbors/leiden in the Next footer), and the troubleshooting table supplies symptom→fix recovery paths. It is not a 5 because there is no explicit verification checkpoint (e.g., confirm out.h5ad exists and reads back after the remote GPU job, or check the embedding shape before downstream clustering); it is not a 3 because the sequence and error-recovery paths are explicit and the operations write new files rather than destructively modifying inputs.

Progressive Disclosure

Everything — the full DE output-column enumeration, the complete Modal remote recipe, and two dense tables — is inlined in a single ~170-line SKILL.md with no reference files, and the one bundle file the body depends on ("exec(open(\"scvi-tools/kernel.py\").read())") is absent from the bundle (no references/, scripts/, or assets/ exist), leaving the central h5ad_safe_obs helper as a dangling reference. It is not a 2 because section structure is clear and well-organized; it is not a 4 because a referenced path is broken and reference-grade material that belongs in separate files sits inline.

Specificity

The description enumerates concrete capabilities — "scVI for a batch-corrected latent space, scANVI for semi-supervised label transfer, and Bayesian differential expression" — plus four concrete use actions ("integrate scRNA-seq batches, embed cells for clustering, transfer annotations from a reference onto a query, or score differentially expressed genes per cluster"), which is comprehensive coverage of what the skill does. It is not a 4 because there are no gaps in capability coverage: every function of the skill (embedding, integration, label transfer, DE) is named as a specific action.

Completeness

It explicitly answers "what" (three named capabilities) and "when" via an explicit trigger clause — "Reach for this skill to integrate scRNA-seq batches, embed cells for clustering, transfer annotations..., or score differentially expressed genes per cluster" — with concrete trigger phrases, and even adds an exclusion boundary ("For spatial deconvolution / mapping use the cell2location, DestVI, or Tangram methods instead"). It is not a 4 because the "when" is fully explicit with specific triggers rather than merely present.

Trigger Term Quality

Natural user phrasings are comprehensively covered with synonyms: "single-cell RNA-seq" / "scRNA-seq", "integrate ... batches", "clustering", "transfer annotations from a reference onto a query" / "label transfer", and "differentially expressed genes", alongside the tool names users would mention. It is not a 4 because both synonyms and task-oriented phrasings (not just tool jargon) are present, matching how a user would naturally phrase these needs.

Distinctiveness Conflict Risk

The skill occupies a clear niche (scvi-tools probabilistic scRNA-seq) with distinct triggers, and the closing boundary sentence actively routes the nearest overlapping use case (spatial deconvolution) to other methods, minimizing wrong-skill triggering. It is not a 4 because even the main overlap risk is explicitly disambiguated rather than left as a minor residual risk.