Conciseness

The body is dense with non-obvious, high-value detail ("Default kernel backend is `None` (reference PyTorch, ~12x slower than paper)", the structseq single-tuple constructor note, the cusolver SVD poison fix) and never explains concepts Claude already knows. It falls short of anchor 5 only through redundancy: the ESMC `<mask>`-token warning appears verbatim in both the body and references/esmc.md, and license/metadata is duplicated between the frontmatter comments and the body. Not 3 — there is no padded or unnecessary explanation.

Actionability

Fully executable, copy-paste-ready guidance throughout: exact pinned install commands, a complete fold() example with annotated paper-faithful parameters, a working SVD monkeypatch, and concrete fallback commands ("fall back to set_kernel_backend(None) + set_chunk_size(64)"). Specific examples cover the common cases (complex folding, MSA input, variant selection, backend tuning), matching anchor 5.

Workflow Clarity

A clear install → usage → throughput → troubleshooting sequence with strong error-recovery guidance for known failure modes (illegal memory access above L≈1400 with fallback, cuequivariance silent-fallback warning, a3m null-byte cleaning, NaN-poisoned cusolver workspace). It misses anchor 5 only because there is no explicit output-validation checkpoint (e.g., sanity-checking the ranked prediction before writing the CIF); it is well above anchor 3, whose checkpoints would be merely implicit.

Progressive Disclosure

Scored against the actual bundle: both referenced files (references/design-hook.md, references/esmc.md) exist, are one level deep with no nesting, and are clearly signaled in context ("see `references/design-hook.md`"; "Full API, mutation scoring, SAE features, contact prediction: see `references/esmc.md`"). The body is an overview with ESMC and design-hook detail appropriately split out — matches anchor 5.

Specificity

Lists multiple concrete actions — "all-atom co-folding", "Single-sequence and MSA modes", "protein, DNA, RNA, ligand (CCD/SMILES), modified residues", plus "masked-LM logits, hidden states, mutation scoring, contact prediction, and the SAE interpretability head" — with comprehensive coverage across both the folding and language-model halves of the release. It exceeds the anchor-4 example because there are no minor gaps in capability coverage.

Completeness

Explicitly answers both what ("all-atom co-folding... protein, DNA, RNA, ligand... masked-LM logits, hidden states, mutation scoring...") and when, via an enumerated "Use this skill when: (1)–(5)" clause with concrete trigger phrases. This matches anchor 5 exactly; anchor 4 would require the 'when' to be less explicit.

Trigger Term Quality

Natural phrases users would say appear throughout ("Predicting complex structures", "Running ESMFold2 with MSA input", "Getting ESMC embeddings or per-residue mutation scores", "Validating designed binders"), but common synonyms like "protein structure prediction", "folding", or "complex prediction" are absent. Good coverage with a few natural terms missing — matches anchor 4, not 3 (many relevant keywords present) or 5 (synonym coverage is incomplete).

Distinctiveness Conflict Risk

The named niche (Biohub ESMFold2/ESMC, HuggingFace org `biohub`) is highly distinct with specific triggers, but trigger (1) "Predicting complex structures with single-sequence input" and (4) "Getting ESMC embeddings" could overlap with closely related structure-prediction or embedding skills. Mostly distinct with minor overlap risk — anchor 4, not 5 (some overlap potential) or 3 (model names and tasks are far more specific than "Works with document files").