Content
63%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A well-structured, highly actionable tool-usage skill with concrete invocations, worked examples, fallback chains, and honest limitation notes. Its weaknesses are redundancy (duplicate tool tables, repeated caveats, restated textbook drug-likeness rules) and the absence of any progressive disclosure — a 350-line monolith whose API reference belongs in a separate file.
Suggestions
Merge the 'Key Tools' and 'Tool Parameter Reference' tables into one table (or move the merged table to references/tool_reference.md) — they list ~25 of the same tools with overlapping parameter columns, and the BindingDB-unavailable caveat is repeated in four sections.
Trim or move the 'Key drug-likeness rules' block and the Lipinski caveats in 'Domain Reasoning' — Claude already knows Lipinski/Veber cutoffs; keep only the skill-specific guidance ('match the profile to the target and route').
Replace placeholder arguments like CANONICAL_SMILES in Phases 2-6 with a single concrete worked example (as Phase 1 does for imatinib) so every phase is copy-paste ready, and add a quick validity check after identity resolution (e.g., confirm a CID was returned before running ADMET tools).
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and tool-focused, but there is measurable waste: the 'Key Tools' table (lines 49-79) and the 'Tool Parameter Reference' table (lines 247-274) duplicate ~25 tools with overlapping parameter columns; BindingDB's flakiness is repeated in at least four places (KEY PRINCIPLES, Phase 3, Fallback Chains, Limitations); and the 'Key drug-likeness rules' block (Lipinski/Veber/lead-like cutoffs) plus the Domain Reasoning paragraph restate standard chemistry knowledge Claude already has. This fits the score-3 anchor (mostly efficient but includes some unnecessary explanation or could be tightened); it is not the score-2 'noticeably verbose, several padded sections' case since most sections carry non-obvious operational facts (list-vs-string SMILES, URL-only returns). | 3 / 5 |
Actionability | Guidance is highly concrete: exact tool names with parameters (e.g., 'ChEMBL_search_activities(molecule_chembl_id="CHEMBL941", pchembl_value__gte=6, limit=50)'), a fully worked Phase 1 example with real values (imatinib -> CID 5291 -> CHEMBL941), a pChEMBL interpretation table, and a fallback-chain table. It falls short of score 5 because later phases use placeholders ('CANONICAL_SMILES', 'SCAFFOLD_SMILES') rather than copy-paste-ready literals, leaving minor gaps; it is well above score 3 since these are real tool invocations, not pseudocode. | 4 / 5 |
Workflow Clarity | Six clearly sequenced phases with an explicit 'ID resolution priority' ordering, per-phase expected outputs ('-> Returns: ...'), and a Fallback Chains table that functions as error-recovery guidance (BindingDB -> ChEMBL, ADMET-AI -> SwissADME, Enamine API 500 -> URL). This matches the score-4 anchor (clear sequence, most checkpoints present, minor validation gaps); it does not reach score 5 because there are no explicit validation steps confirming lookups succeeded before proceeding (e.g., verifying a CID resolved before running ADMET), and the workflows are read-only so the destructive/batch cap does not apply. | 4 / 5 |
Progressive Disclosure | The skill is a single 350-line SKILL.md with no bundle files (references/, scripts/, assets/ do not exist), and roughly 90 lines are API-reference tables ('Key Tools', 'Tool Parameter Reference') that the score-2/3 anchors treat as content that belongs in a separate reference file. Section headers and horizontal rules give it real structure, so it sits at the score-3 anchor (some structure, content that should be separate is inline) rather than score 2, and short of score 4 since there is no split at all despite the file being long. | 3 / 5 |
Total | 14 / 20 Passed |