Conciseness

The body is dense with zero concept-explanation padding — every section is a pinned fact, gotcha, or executable snippet, and it assumes Claude's competence throughout. It stops short of the score-5 'every token earns its place' anchor because the fold hyperparameters (10 loops, 68 steps, 5 samples) are stated three times: in the fold() code comments, again in the 'Paper-faithful FoldBench settings' paragraph, and a third time in the paper-matched configuration table.

Actionability

Fully executable throughout: a copy-paste install with exact version pins and commit hashes, a complete fold() example from imports to mmCIF export, a working safe-SVD monkeypatch, and a runnable MSA input example. Specific examples cover the common cases (complex, monomer, MSA, variants), matching the score-5 anchor exactly.

Workflow Clarity

The install → load → set_kernel_backend → fold → rank-by-ipTM sequence is clear, and error-recovery loops are unusually thorough (L>1400 illegal-memory-access fallback, the SVD-poison monkeypatch, a3m null-byte cleaning, backend selection rules). It sits at score 4 rather than 5 only because there is no explicit validation checkpoint for the core flow — e.g., a quick import sanity check after the fragile pinned install, or a quality gate on the ranked prediction before writing it out.

Progressive Disclosure

The body is a well-sectioned overview with clearly signaled, one-level-deep references — 'gradient-guided design — see references/design-hook.md' and 'Full API, mutation scoring, SAE features, contact prediction: see references/esmc.md' — and both files exist in the bundle. The two deep-dive topics (design hook, ESMC API) are appropriately split out while day-to-day content stays inline, matching the score-5 anchor.

Specificity

The description enumerates concrete capabilities: 'all-atom co-folding', 'Single-sequence and MSA modes; protein, DNA, RNA, ligand (CCD/SMILES), modified residues', and for ESMC 'masked-LM logits, hidden states, mutation scoring, contact prediction, and the SAE interpretability head'. This matches the anchor for multiple specific concrete actions with comprehensive coverage; it goes beyond the score-4 anchor's 'minor gaps in coverage'.

Completeness

It explicitly answers both questions: the 'what' is a detailed multi-clause capability list, and the 'when' is an explicit 'Use this skill when: (1)...(5)' enumeration with concrete trigger scenarios. This is a direct match to the score-5 anchor; score 4 would require the 'when' to be less explicit than it is.

Trigger Term Quality

Natural trigger phrases are present ('Predicting complex structures', 'Validating designed binders', 'Running ESMFold2 with MSA input', 'Getting ESMC embeddings or per-residue mutation scores'), giving good keyword coverage. It falls short of the score-5 anchor because common variations users would say — plain 'protein folding' / 'structure prediction' phrasing and file extensions like .a3m, .cif, or .pdb — are absent.

Distinctiveness Conflict Risk

The niche is unambiguous — named models (ESMFold2/ESMFold2-Fast/ESMC sizes), the paper citation, github.com/Biohub/esm, and the HF org 'biohub' — so it would not plausibly trigger for a different skill. It matches the clear-niche anchor with minimal conflict risk.