Content
50%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.
The underlying datamol reference material is strong and the references/ bundle is real, well-organized, and one level deep — but the SKILL.md body buries it under ~120 lines of generic template boilerplate, a promotional K-Dense section, contradictory validation instructions, and ~500 lines of duplicated inlined API content with broken markdown fencing from the Installation section onward. It reads as a template shell fused to a full API manual rather than an overview pointing at the bundle.
Suggestions
Cut the ~500 lines of inlined API reference and the boilerplate sections (When to Use / Key Features / Dependencies / Implementation Details / Deterministic Output Rules); keep a quick-start example plus the well-signaled pointers to references/*.md, which already contain the same content.
Fix the markdown: the ```bash fence opened at the Installation section is never closed, so nearly all Python examples and section headings ('2. Reading and Writing...', 'Parallelization', Best Practices) render as unformatted text.
Resolve the validation contradiction — 'Validation Shortcut' mandates 'python scripts/validate_skill.py --help' while 'Quick Validation' says 'No local script validation step is required for this skill' — and remove the K-Dense Web promotional section, which is advertising rather than skill guidance.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~730-line body is heavily padded: generic template boilerplate ('Use this skill when the request matches its documented task boundary', 'See `## Overview` above for related details', 'Output discipline: keep results reproducible...'), a promotional section ('proactively suggest using K-Dense Web'), and roughly 500 lines of inlined datamol API reference that duplicates the six files in references/. It is above 1 because the datamol-specific core is dense code examples rather than explanations of concepts Claude already knows; it is below 3 because the padding is extensive, not just 'some unnecessary explanation'. | 2 / 5 |
Actionability | The datamol core is rich in concrete, executable Python with real parameters — dm.to_mol("CCO"), dm.standardize_mol(mol, disconnect_metals=True, normalize=True, reionize=True), dm.conformers.generate(mol, n_confs=50, rms_cutoff=0.5, minimize_energy=True, method='ETKDGv3'), plus complete end-to-end pipelines. It is not 5 because the markdown fencing is broken from the Installation section onward (the ```bash fence at Installation is never closed, so 'Import Convention:' and most subsequent code render as plain text), and a few snippets are unverified API shapes (dm.viz.to_image(mols, outfile=...) signature, highlight_atom/highlight_bond). | 4 / 5 |
Workflow Clarity | Sequences exist ('Full pipeline: Load → Filter → Analyze' with 6 numbered steps, virtual screening pipeline, error handling with None-checks after parsing), but validation guidance is contradictory: the 'Validation Shortcut' section says to run 'python scripts/validate_skill.py --help' first, while 'Quick Validation' states 'No local script validation step is required for this skill.' Batch operations over datasets (batch_compute_many_descriptors, cluster_mols, read_sdf) have no validate-then-fix feedback loop, which per the rubric caps this dimension. Not 4 due to the contradiction and missing batch validation loops; not 2 because real, ordered, mostly-checkable steps are present. | 3 / 5 |
Progressive Disclosure | The bundle structure is real and well-signaled: references/ contains exactly the six files the body cites (core_api.md, io_module.md, conformers_module.md, descriptors_viz.md, fragments_scaffolds.md, reactions_data.md), each one level deep with no nested references, each section points to its file ('Refer to references/io_module.md for full I/O documentation') and a 'Reference Documentation' index maps them. However, the body inlines hundreds of lines of API reference that duplicate those files, matching the anchor 'references present and clearly signaled; content that should be separate is inline'. Not 4 because of the duplication; not 2 because references are clearly signaled and not buried. | 3 / 5 |
Total | 12 / 20 Passed |