Conciseness

The body is mostly efficient, code-forward, and assumes competence, but includes repeated and off-scope content: the backed-mode try/finally close pattern appears four times (Reading data, Core Capabilities, Best Practices, Troubleshooting), the scanpy 10X read example appears three times, and a full scanpy preprocessing pipeline (normalize/log1p/PCA/UMAP/Leiden) is inlined despite the description itself routing analysis workflows to the scanpy skill. Not 4 because the duplication and the off-scope scanpy pipeline are more than minor trimmable instances; not 2 because most sections are lean, executable, and free of concepts-Claude-already-knows padding.

Actionability

Nearly all guidance is copy-paste-ready executable code covering the common cases (creation, read/write, backed mode, concat, subsetting, QC filtering, index alignment with an explicit wrong/correct pair). Not 5 because of minor gaps: `process(adata_subset)`/`process(chunk)` and undefined `adata1/adata2/adata3` placeholders are illustrative rather than runnable, and the TileDB-SOMA section advises verification without showing any command.

Workflow Clarity

Multi-step workflows are clearly sequenced (numbered single-cell RNA-seq pipeline, batch combination, chunked large-dataset processing) with real checkpoints: `assert adata.obs['batch'].notna().all()`, "Write those edits to a new file and reopen it to verify they survived", try/finally file closes, and the layers['counts'] provenance snapshot before normalization. Not 5 because there is no explicit validate→fix→retry feedback loop for the fragile write/reopen operations; not 3 because validation steps are present and explicit, not merely implicit.

Progressive Disclosure

The body is a genuine overview: each capability section is a short summary plus a clearly signaled one-level-deep pointer ("**See**: references/data_structure.md" with an itemized contents list), and all five referenced files exist in the bundle. Verified that reference files contain only external documentation URLs, not nested .md references, so navigation is one level deep and easy.

Specificity

"Handles annotated matrices in single-cell analysis, .h5ad and Zarr files, and integration with the scverse ecosystem" names the domain, concrete file formats (.h5ad, Zarr), and a specific ecosystem integration, giving several specific anchors. Not 5 because the only verb is the generic "Handles" — unlike the anchor's multiple distinct actions (extract, fill, merge, convert), no individual operations (read/write/concat/subset) are named; not 3 because the format and ecosystem coverage goes beyond 1-2 bare actions.

Completeness

The "what" is clear (annotated matrices, .h5ad/Zarr files, scverse integration), and "when" is addressed via explicit routing: "This is the data format skill—for analysis workflows use scanpy; for probabilistic models use scvi-tools; for population-scale queries use cellxgene-census." Not 5 because the trigger guidance is phrased negatively (when to use other skills) rather than as a direct positive "Use when creating, reading, or writing AnnData objects" clause with concrete trigger phrases; not 3 because the routing is explicit and actionable, well beyond weakly implied.

Trigger Term Quality

Good natural-term coverage: "single-cell analysis", ".h5ad", "Zarr", "scverse", plus sibling tool names (scanpy, scvi-tools, cellxgene-census) that users actually say. Not 5 because the description never mentions "AnnData"/"adata" itself — the package name and object name users would most naturally type — nor synonyms like "h5ad" without the dot.

Distinctiveness Conflict Risk

The description carves out a clear niche (the data-format layer) and explicitly disambiguates against the three closest sibling skills (scanpy for analysis, scvi-tools for models, cellxgene-census for population queries), matching the anchor for a clear niche with distinct triggers and minimal conflict risk.