Conciseness

The body is dense and assumption-respecting throughout — a prerequisites table, exact install/download commands, a query-JSON spec with a validation note ('extra: forbid — unknown keys reject'), a flag table, an output tree, and a compact troubleshooting table. No section explains concepts Claude already knows (no 'what is a protein' or library tutorials); every line carries non-obvious operational detail such as the DeepSpeed-vs-cuEquivariance kernel swap and the eager-imported boto3 gotcha.

Actionability

Everything is copy-paste executable: 'pip install \'openfold3[cuequivariance]==0.4.1\'', the 'huggingface-cli download' command with a pinned checkpoint path, a complete 'run_openfold predict' invocation, a valid query JSON example, and verification commands ('grep -E \'Successful|Failed\' out/summary.txt', 'find out -name '*_model.cif' | wc -l') with the expected-count formula. The troubleshooting table maps exact error strings to exact fixes ('pip install nvidia-cutlass', 'apt-get install libxrender1 libxext6 libsm6').

Workflow Clarity

The flow is clearly sequenced — prerequisites, install, weights, run, query format, output interpretation — and closes with an explicit 'Verify' section containing feedback criteria ('Successful Queries: N matching your input count', the queries × seeds × samples count check) plus quality thresholds (avg_plddt, ptm/iptm, has_clash) and a diagnosis/fix table for failure recovery. Not anchor 4 because validation checkpoints are explicit, not merely implied.

Progressive Disclosure

Sections are well-organized and the SKILL.md works standalone (no references/ scripts/ assets/ bundle exists, so nothing is buried or nested). It stops short of anchor 5 because at ~150 lines the query-format reference table and the 8-row troubleshooting table are prime candidates for one-level-deep reference files (e.g. TROUBLESHOOTING.md) that would keep the top-level file a lean overview; the guideline's 5-with-sections-alone exception applies only to skills under 50 lines.

Specificity

The description names concrete actions — 'Structure prediction' and 'predicting protein/nucleic-acid/ligand complex structures' — with comprehensive scope across molecule types and naming the exact tool ('OpenFold3, an open-weights PyTorch reproduction of AlphaFold3 from the AlQuraishi Lab'). It sits above anchor 3 (only 1-2 actions, not comprehensive) because the prediction scope is fully enumerated, but below anchor 5 because only one action verb is listed rather than multiple distinct actions.

Completeness

Both parts are explicit: the 'what' ('Structure prediction using OpenFold3, an open-weights PyTorch reproduction of AlphaFold3') and the 'when' ('Use this skill when predicting protein/nucleic-acid/ligand complex structures...') with concrete trigger phrases. It is not anchor 4 because the 'when' clause is not merely implied or generic — it directly names the triggering task.

Trigger Term Quality

Strong natural keyword coverage: 'OpenFold3', 'AlphaFold3', 'AF3', 'structure prediction', 'protein/nucleic-acid/ligand', 'complex structures' — terms a user asking for structure prediction would genuinely say. It falls short of anchor 5 because common variations like 'protein folding', 'fold a protein', or output-format terms (mmCIF/.cif) are absent.

Distinctiveness Conflict Risk

Clear niche anchored to uniquely named tools (OpenFold3, AlphaFold3) and a specific task (protein/nucleic-acid/ligand complex structure prediction), making accidental triggering by unrelated skills essentially impossible. Nothing in the wording overlaps with generic domains.