Conciseness

The body is mostly high-value operational knowledge (MAD gating rules, tool parameter-validation quirks, pipeline ordering), but it includes unnecessary explanation of concepts Claude already knows — e.g., the 'WHY each metric flags a problem' section explaining membrane rupture enriching mitochondrial transcripts, the full Scanpy-vs-Seurat equivalents table, and generic statements like 'A database-verified answer is always more reliable than a guess'. At ~430 lines with a ~70-line inline ToolUniverse tool catalog, it could be meaningfully tightened. This matches the score-3 anchor ('mostly efficient but includes some unnecessary explanation or could be tightened') better than score 4, where over-explanation would be only minor.

Actionability

Guidance is fully executable throughout: copy-paste-ready data-loading code, QC metric computation code, a complete end-to-end pipeline block (read -> QC -> normalize -> HVG -> PCA -> cluster -> annotate -> DE), statistical test snippets with exact function signatures, and concrete commands like `python scripts/scrna_qc.py data.h5ad --doublets` and a fully-parameterized `tu run run_deseq2_analysis` invocation. Even failure paths are executable ('print the install plan and stop'). This matches the score-5 anchor; code is real, not pseudocode, so score 3/4 anchors do not apply.

Workflow Clarity

The sequence is clear and well-ordered: RULE ZERO (check pre-computed results before re-analysis), a workflow decision tree routing each question type to a reference, an explicit QC-before-downstream ordering with per-step removal reporting, and a troubleshooting table for error recovery. Checkpoints are mostly present (distribution visualization before cutoffs, --install-plan preflight, 'never report cutoffs you did not actually run'), but there is no explicit validate-output-then-retry loop after the pipeline (e.g., re-validate clustering after QC removal), which keeps it at score 4 rather than 5. It is clearly above score 3, where the sequence would exist but checkpoints would be missing.

Progressive Disclosure

Structure is generally good — a decision tree pointing to eight one-level-deep references/*.md files, all of which exist — but the body links to `analysis_patterns.md` three times (the decision tree's 'Pattern 1' and 'Pattern 2' for the most common DE and correlation patterns, and the Reference Documentation section) and that file does not exist in the bundle, leaving a dead link on the primary navigation path. Additionally, the ~70-line ToolUniverse tool catalog is inlined in SKILL.md where it belongs in a reference file. This fits the score-3 anchor ('could be better organized; content that should be separate is inline') better than score 4's 'minor organization gaps', since a missing referenced file is a navigation failure, not a minor gap.

Specificity

The description lists multiple specific concrete actions — 'h5ad data loading', 'quality control and QC gating', 'normalization', 'dimensionality reduction (PCA, UMAP, t-SNE)', 'clustering (Leiden, Louvain)', 'marker gene identification', 'cell-type annotation', 'pseudotime/trajectory analysis' — and even names the exact QC metrics (n_genes_by_counts, pct_counts_mt, MAD-based thresholds) and tools (Scrublet/scDblFinder, SoupX). Coverage is comprehensive across the single-cell workflow, matching the score-5 anchor; the score-4 anchor ('minor gaps in coverage') fits less well since both common and advanced steps are enumerated.

Completeness

Both 'what' and 'when' are explicitly answered: the capability list answers what the skill does, and 'Use for any scRNA-seq workflow, including deciding which cells to filter, flag, or investigate before downstream analysis' is an explicit when-to-use clause with a concrete trigger phrase. This matches the score-5 anchor; it is well above score 4, where the 'when' would only be present but less specific.

Trigger Term Quality

Natural user phrasings are comprehensively covered with synonyms and formats: 'single-cell RNA-seq' and 'scRNA-seq', 'h5ad', 'QC', 'clustering', 'PCA, UMAP, t-SNE', 'marker gene', 'cell-type annotation', 'pseudotime', 'trajectory', 'deciding which cells to filter, flag, or investigate'. Tool names users would mention (scanpy, anndata, Scrublet, SoupX) are also present. This matches the score-5 anchor ('synonyms and file extensions'); the score-4 anchor applies only when natural terms are missing, and none of the common scRNA-seq vocabulary is absent.

Distinctiveness Conflict Risk

The niche is clear and distinct — single-cell RNA-seq analysis with scanpy/anndata — with specific metric and algorithm names that few other skills would claim. However, 'Use for any scRNA-seq workflow' is a deliberately broad catch that has minor overlap risk with closely related sibling skills (e.g., deep-learning scRNA integration/batch correction covered by tooluniverse-scvi-tools). This fits the score-4 anchor ('mostly distinct; minor overlap risk with closely related skills') better than score 5's 'minimal conflict risk', but clearly above score 3, which would require only vague domain identification.