Conciseness

The body is ~420 lines and, while mostly domain-specific knowledge Claude does not have (QC gating rationale, MAD thresholds, ToolUniverse tool names), it includes sections that re-teach what Claude already knows: the scipy 'Statistical Tests (Quick Reference)' (pearsonr, ttest_ind, f_oneway, multipletests), the 'Scanpy vs Seurat Equivalents' table, and a ~40-line inline ToolUniverse tool catalog. This is 'mostly efficient but includes some unnecessary explanation or could be tightened' — not severely padded enough for 2, but clearly not the lean 4-5 anchors.

Actionability

Guidance is fully executable throughout: copy-paste data loading with orientation handling, a runnable QC metrics block with the 'percent_top=None' gotcha, 'python scripts/scrna_qc.py data.h5ad --doublets', 'python scripts/scrna_qc.py --install-plan', a complete scanpy pipeline, working harmonypy code, and exact 'tu run ...' JSON command lines. This matches the 5 anchor: copy-paste-ready commands covering the common cases, with no pseudocode.

Workflow Clarity

The workflow decision tree, RULE ZERO pre-computed-results check, QC-before-downstream ordering, and the QC interpretation table give a clear sequence with most checkpoints present ('Always visualize distributions first', 'never report cutoffs you did not actually run', the helper 'reports per-step removals', and the honest-execution install-plan guard for the batch filtering operation). It sits below 5 only because there is no explicit validate-then-re-filter feedback loop and the 'Complete Pipeline (Quick Reference)' contradicts the MAD guidance by hardcoding 'pct_counts_mt < 20' — minor validation/consistency gaps, hence 4 not 3 since checkpoints for the destructive filtering steps are explicitly present.

Progressive Disclosure

The bundle has 8 reference files, all clearly listed with one-line descriptions under 'Reference Documentation' and cross-linked from the decision tree and QC section — good signaling. However, 'analysis_patterns.md' is referenced three times ('See: analysis_patterns.md "Pattern 1"', 'Pattern 2', and in Reference Documentation) but does not exist anywhere in the bundle, and sizable catalogs (ToolUniverse tools, Seurat equivalents, stats basics) are inlined in SKILL.md rather than split out. A dangling reference plus inlined reference-material keeps this at 'some structure but could be better organized' rather than the 4 anchor's 'minor organization gaps'.

Specificity

The description lists multiple concrete actions with comprehensive coverage — 'h5ad data loading, scRNA-seq quality control and QC gating (n_genes_by_counts, total_counts, mitochondrial percent / pct_counts_mt... doublet detection with Scrublet/scDblFinder, ambient RNA / SoupX awareness, empty-droplet filtering, MAD-based thresholds), normalization, dimensionality reduction (PCA, UMAP, t-SNE), clustering (Leiden, Louvain), marker gene identification, cell-type annotation, pseudotime/trajectory analysis'. This matches the 5 anchor ('lists multiple specific concrete actions; comprehensive coverage') — it names exact metrics, tools, and algorithms rather than generic verbs, and nothing is vague enough to drop to 4.

Completeness

It explicitly answers both: what — the full capability list above; and when — 'Use for any scRNA-seq workflow, including deciding which cells to filter, flag, or investigate before downstream analysis'. This matches the 5 anchor (clear concrete 'what' plus explicit 'Use when' trigger phrases) exactly, and the 'when' clause is specific rather than weakly implied.

Trigger Term Quality

Strong natural keyword coverage: 'Single-cell RNA-seq', 'scRNA-seq', 'h5ad', 'QC', 'UMAP', 't-SNE', 'Leiden', 'clustering', 'marker gene', 'cell-type annotation', 'pseudotime'. A few natural terms users commonly say are missing — '10X', 'Cell Ranger', 'count matrix', 'transcriptomics' — so it falls just under the 5 anchor's 'comprehensive coverage including synonyms and file extensions'.

Distinctiveness Conflict Risk

It carves out a clear niche (single-cell RNA-seq with scanpy/anndata) with distinct triggers (scRNA-seq, h5ad, Leiden, Scrublet) that would not fire for bulk RNA-seq, enrichment, or variant skills, matching the 5 anchor 'clear niche with distinct triggers; minimal conflict risk'. Third-person voice is used throughout, so no specificity penalty applies.