Conciseness

The ~70-line body is lean and every non-obvious fact earns its place — the YAML-overwrites-CLI gotcha ("replaces every key it finds, so passing --samples_per_complex 40... is silently ignored"), the silent ~11-minute SO(3) precompute with its 32 GB RAM requirement, and the --ligand prefix-matching trap. No concepts Claude already knows are re-explained. Matches the 'every token earns its place' anchor.

Actionability

The main invocation is copy-paste ready with real flags (python3 -m inference --config default_inference_args.yaml --protein_path target.pdb --ligand_description "COc1ccc(C#N)cc1"), the batch path specifies the exact four CSV columns, the output convention (rank{N}_confidence{score}.sdf) is documented, and the error table gives concrete fixes (sed the setrlimit constant to min(64000, rlimit[1])). Matches the fully-executable top anchor.

Workflow Clarity

The single-complex path is unambiguous and the "Errors worth recognizing" table provides explicit error-to-fix recovery loops, but there is no post-run verification checkpoint for the batch/fragment-library path (e.g., confirm each complex produced a rank1.sdf before comparing ligands). This sits between 'most checkpoints present' (4) and 'explicit validation steps with feedback loops' (5).

Progressive Disclosure

The body keeps the single-complex path inline and correctly splits batch docking and the ESMFold sequence-only path into references/workflows.md (a real file, one level deep, clearly signaled in context: "that path and a larger-library screening recipe are in references/workflows.md"). Structure is well organized with earned section headings; matches the top anchor.

Specificity

The description lists multiple concrete, distinct actions — "Predict small-molecule binding poses", "places a ligand into a protein pocket without a predefined search box", "ranks the samples with a learned confidence model", "dock a SMILES or SDF against a PDB", "generate ranked 3D poses for a small fragment library", "get a starting pose for downstream rescoring" — covering the tool's use modes comprehensively, which matches the top anchor rather than the 'minor gaps' level below.

Completeness

It explicitly answers both what ("blind diffusion docking that places a ligand into a protein pocket... and ranks the samples with a learned confidence model") and when ("Reach for this skill to dock a SMILES or SDF against a PDB, to generate ranked 3D poses for a small fragment library, or to get a starting pose for downstream rescoring"), with concrete trigger phrases plus a scope boundary ("predicts geometry, not affinity").

Trigger Term Quality

Natural trigger terms a docking user would actually say are covered including formats and synonyms: "dock", "SMILES", "SDF", "PDB", "binding poses", "fragment library", "rescoring". This matches the comprehensive-coverage anchor; score 4 would require missing natural terms, and none are obviously absent for this niche.

Distinctiveness Conflict Risk

It names the specific tool (DiffDock-L, Corso et al., github.com/gcorso/DiffDock) and its distinct niche (blind docking without a predefined search box, SMILES/SDF against a PDB), giving minimal overlap risk with sibling modeling skills. Not 4: the niche and triggers are clearly distinct rather than having minor overlap.