Content
87%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The body is dense, executable, and well-structured: runnable commands, exact schemas, and high-value gotchas with copy-paste fixes, all kept lean. The only gap is an explicit post-run validation checkpoint for the batch path, which leaves workflow clarity just short of full marks.
Suggestions
Add an explicit validation step for batch runs (e.g. after `--protein_ligand_csv`, check that each complex produced a rank1.sdf and that no row was silently treated as ligand-less) so the batch workflow has a clear validate→fix→retry checkpoint.
Note how to confirm the SO(3) precompute completed vs. a genuine hang beyond 'do not assume a hang means a crash' — e.g. the stderr/log line or timestamp to look for — to make that checkpoint actionable rather than cautionary.
Make the rescoring hand-off at the end slightly more actionable by naming the concrete command or reference section for boltz's affinity head, rather than only describing it.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Lean throughout with no padding of concepts Claude already knows (no 'what is a PDB/SMILES'); every section — the run command, YAML-overwrite caveat, silent-precompute gotcha, and errors table — earns its tokens. The only time-sensitive stamp (license verification 2026-06-30) lives in a YAML comment, not the body, so it does not penalize conciseness. | 3 / 3 |
Actionability | Provides a fully executable `python3 -m inference` command with real flags, the exact four-column CSV schema, the exact output filename pattern `rank{N}_confidence{score}.sdf`, and copy-paste fixes in the errors table (e.g. sed the setrlimit constant to `min(64000, rlimit[1])`). | 3 / 3 |
Workflow Clarity | The single-complex run is unambiguous and the gotchas act as implicit checkpoints ('do not assume a hang means a crash', the memory fix, the YAML-overwrite warning), but for the batch/CSV path there is no explicit validate→fix→retry checkpoint on outputs, so it sits at score 2 rather than 3. | 2 / 3 |
Progressive Disclosure | The body is an overview that defers the two secondary entry points (batch/library screening, sequence-only ESMFold) to references/workflows.md, which is a real, one-level-deep, clearly signaled reference with no nested indirection. | 3 / 3 |
Total | 11 / 12 Passed |