Lean and efficient — no padding about what single-cell biology or scGPT is; every section (prerequisites, run, output, remote compute, gotchas, troubleshooting) earns its place and assumes Claude's competence. It is not the level-2 case because there is no unnecessary explanation that could be trimmed.

Provides fully executable, copy-paste-ready code with specific values — GeneVocab.from_file on vocab.json, embed_data with gene_col and use_fast_transformer, and a complete submit_job call with real parameters (timeout 1800, 24 GB VRAM, named inputs/outputs).

The remote-compute process is clearly sequenced with an explicit checkpoint: submit_job → 'cell ends here — kernel never blocks' → wait_for_notification → 'When the compute_done notification arrives, act on its payload' → re-enter kernel to attach_job/close, with a troubleshooting table providing error→fix feedback loops.

No bundle files exist, and the single SKILL.md is organized into well-signaled sections (Prerequisites, How to run, Output format, Remote compute, Gotchas, Troubleshooting) with no nested/deep references — well-organized sections suffice for this skill, per the simple-skills guidance.

Lists multiple concrete actions — 'Producing cell embeddings from an AnnData for clustering/integration', 'cell-type annotation', 'Gene-level representation for perturbation/GRN tasks' — matching the score-3 anchor of several specific concrete actions.

Explicitly answers both 'what' ('Embed and annotate single-cell expression data with scGPT') and 'when' ('Use this skill when: (1)... (2)... (3)...'), matching the score-3 anchor with an explicit trigger clause.

Uses natural domain terms a single-cell practitioner would say — 'single-cell expression data', 'cell embeddings', 'AnnData', 'clustering/integration', 'cell-type annotation', 'perturbation/GRN' — giving good coverage rather than jargon-only phrasing.

Occupies a clear niche (scGPT single-cell foundation model) and explicitly routes the overlapping case elsewhere ('For probabilistic single-cell models (scVI etc.), use the scvi-tools library'), making mis-triggering unlikely.