Conciseness

Largely lean — terse tables, executable code, and scvi-specific caveats Claude would not know — but a few passages over-explain (e.g. 'PyTorch MPS and MLX are different backends', the detailed arXiv-fetching instructions in the citation section), sitting between 'efficient with minor trims' and 'every token earns its place'.

Actionability

The RNA workflow, DE example, and install commands are fully executable with assertion checkpoints, but specialized-model paths (SCANVI annotation, TOTALVI, reference mapping) are pointer-only with no inline runnable example, matching 'concrete code with minor gaps' rather than full coverage of common cases.

Workflow Clarity

The five-step Workflow section has explicit validation (step 4: 'Inspect validation history, held-out fit, seed stability, batch mixing and retention of known biology'), the code includes assertion checkpoints, and there are concrete feedback loops ('If filtering removes all expression from a cell, remove that cell or revise feature selection'; 'If the training object changes, register it again') plus a pre-flight check ('Check a small CPU fit and finite outputs... before a long run').

Progressive Disclosure

The body is a genuine overview: primary workflow inline, with all eight referenced files verified to exist, one level deep, clearly signaled by a bulleted per-scope list ('Read the corresponding references before using a specialized model') and targeted inline pointers to workflows, DE, and theory references.

Specificity

Lists multiple concrete, named capabilities — 'scVI batch integration, scANVI annotation, totalVI CITE-seq, MultiVI RNA/ATAC integration, and posterior differential expression' — giving comprehensive coverage of the skill's domain, matching the top anchor rather than the 'minor gaps' anchor below.

Completeness

Explicitly answers both what ('Fits probabilistic models for single-cell omics, including...') and when ('Use for generative modeling, reference mapping, multimodal analysis, or model-based uncertainty'), with concrete trigger phrases and even a boundary clause ('use scanpy for standard preprocessing and exploratory analysis').

Trigger Term Quality

Good natural keyword coverage ('batch integration', 'annotation', 'CITE-seq', 'differential expression', 'reference mapping', 'multimodal analysis'), but common variations users would say such as 'scRNA-seq', 'batch correction', or file extensions like 'h5ad' are missing, so it falls between the 'few natural terms missing' and 'comprehensive including synonyms/extensions' anchors.

Distinctiveness Conflict Risk

Clear niche (model-based single-cell analysis) with distinct model-class triggers and an explicit disambiguation against the nearest competing skill (scanpy), so conflict risk is minimal.