Conciseness

The body is lean and operational throughout — no explaining what AnnData, foundation models, or clustering are; it jumps straight to loading vocab, calling embed_data with annotated parameters, and a compact prerequisites table. The one longer narrative passage (remote-compute handle semantics: "`.close()` lives on the handle, not on the job object") is harness-specific knowledge Claude would not know, so it earns its tokens. Every section maps to the score-5 anchor ("assumes Claude's competence; every token earns its place"); it is not score 4 because there is no identifiable padding to trim.

Actionability

Core flows are copy-paste ready: GeneVocab.from_file with a verification print, a complete embed_data call with typed arguments, a full submit_job invocation, and the attach/close sequence. The few placeholders ("/path/to/scgpt-human", "environment=...") are explicitly justified — each is annotated with where to obtain the real value ("env name from compute_details", checkpoint path from compute_details), and troubleshooting gives an exact fix command (`name.lower().replace('-', '_')`). This matches the score-5 anchor ("specific examples cover the common cases") rather than 4's "minor gaps", since no example is pseudocode or missing key details.

Workflow Clarity

The remote-compute workflow is clearly sequenced: compute_details → host.compute.create → submit_job → wait_for_notification → save_artifacts → attach_job for the full result, with checkpoints like `print(len(gv)) # 60697` and a symptom→fix troubleshooting table (e.g. "Nearly all genes dropped → Wrong gene_col") serving as feedback loops. It falls short of 5 because the batch embed job has no explicit output-validation step (e.g. verifying embedded.h5ad retains the input cell count) — a minor validation gap per the score-4 anchor, not the missing-validation case of 3, since failure symptoms and recovery paths are documented.

Progressive Disclosure

Sections are well-organized (Prerequisites, How to run, Output format, Remote compute, Gotchas, Troubleshooting, Next) and deep orchestration detail is correctly deferred one level to sibling skills ("See the `remote-compute-ssh` / `remote-compute-modal` skill"), with no nested references. It does not reach 5 because the ~120-line body inlines everything (troubleshooting and remote-compute detail could live in reference files), exceeding the under-50-lines case the rubric exempts; structure is good with minor organization gaps, matching the score-4 anchor.

Specificity

The description enumerates the model's full task surface with concrete actions: "Embed and annotate single-cell expression data", "Producing cell embeddings from an AnnData for clustering/integration", "Zero-shot or fine-tuned cell-type annotation", and "Gene-level representation for perturbation/GRN tasks". This matches the anchor for multiple specific concrete actions with comprehensive coverage; it exceeds the score-4 anchor ("several specific actions; minor gaps") because all three major scGPT use-cases are named, and nothing here is vague enough to fall to 3.

Completeness

It explicitly answers both questions: the "what" ("Embed and annotate single-cell expression data with scGPT, a foundation model for single-cell biology") and the "when" via an explicit "Use this skill when:" clause with three concrete trigger conditions, matching the score-5 anchor. It is not score 4 ("'when' could be more explicit") because the trigger list is fully enumerated, and the added scVI routing line further sharpens when-not-to-use.

Trigger Term Quality

Good natural keyword coverage: "single-cell", "cell embeddings", "cell-type annotation", "clustering", "integration", "perturbation", "GRN", "AnnData", "zero-shot", "fine-tuned" — terms a bioinformatics user would plausibly say. It falls short of the score-5 anchor (comprehensive synonyms/extensions, e.g. "PDF files, PDFs, .pdf") because common variants like "scRNA-seq", "batch correction", or ".h5ad" are absent, but it is clearly above the score-3 anchor ("some relevant keywords but missing common variations") since most natural phrasings are present.

Distinctiveness Conflict Risk

It occupies a clear niche (the scGPT foundation model) and actively routes away a neighboring domain: "For probabilistic single-cell models (scVI etc.), use the scvi-tools library". Triggers are tied to scGPT-specific operations, giving minimal conflict risk per the score-5 anchor; the only theoretical overlap is with generic scanpy-clustering skills, which the embedding/annotation framing keeps minor — not enough to drop to 4.