The body is lean — prerequisites/install/scoring/generation are given as compact tables and executable snippets, and prose only adds non-obvious domain guidance (e.g. 'More negative ⇒ less likely', HF_HOME/HF_HUB_OFFLINE rationale). It assumes Claude's competence and avoids explaining what DNA or libraries are, matching the 'every token earns its place' anchor.

Provides copy-paste-ready executable code ('from evo2 import Evo2', 'model.score_sequences(seqs)', 'model.generate(...)') plus concrete install, model-selection, and remote-compute submission snippets with specific VRAM/timing numbers. Fully instructs rather than describes; matches the level-3 anchor.

Sequences are explicit: a decision tree routes to the right model, and the remote-compute flow is step-sequenced (submit_job → wait_for_notification → save_artifacts → attach_job). The troubleshooting table supplies a symptom→cause→fix error-recovery feedback loop. Not level 2 because checkpoints (notification, error recovery) are present rather than implicit; operations are non-destructive inference so strict validate-before-proceed is less critical.

Well-organized into clear headered sections, but it is a monolithic ~125-line file with no bundle references (references/scripts/assets absent); detailed material such as the remote-compute orchestration and the troubleshooting/performance tables is inline where separate reference files would let SKILL.md act as an overview. Fits the level-2 anchor ('content that should be separate is inline') rather than level 3, which expects well-signaled one-level-deep references and appropriately split content.

Lists multiple concrete actions — 'Score, embed, and generate DNA sequences' — elaborated across four enumerated capabilities (likelihoods, embedding, generation, cross-species scoring). Not vague; matches the 'lists multiple specific concrete actions' anchor rather than the domain-only level 2.

Explicitly answers both what ('Score, embed, and generate DNA sequences with Evo 2') and when ('Use this skill when: (1)…(4)'). Matches the level-3 anchor with explicit triggers; not level 2 since the 'when' is stated, not merely implied.

Uses natural domain terms a genomics user would say — 'variant effect scoring', 'embedding genomic windows', 'generating DNA', 'scoring regulatory or coding regions' — with good coverage. Not reduced to 2 because it avoids pure jargon and captures the common phrasings users would actually request.

Clear DNA language-model niche with triggers unlikely to fire for non-DNA skills; the body's decision tree even distinguishes it from borzoi and esm2. Not level 2 because the scope is tightly scoped to DNA likelihood/embedding/generation rather than overlapping 'document/code' generics.