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 dense, experience-driven skill whose strongest material — the deterministic density script, question-to-metric mappings, and exact ToolUniverse parameter quirks — is highly actionable. Its weaknesses are redundancy (the rows-vs-sites rule stated three times), a stale date-stamped API note, and a reference layer that is named but entirely absent from the bundle.
Suggestions
Merge the three overlapping sections ("TOP-OF-MIND RULE", the PRIMARY SCRIPT mapping table, and "Distinguish rows vs unique sites") into one section with a single question-to-metric table, cutting roughly a third of the body without losing information.
Add the referenced files to the bundle (or remove the references): CODE_REFERENCE.md, TOOLS_REFERENCE.md, ANALYSIS_PROCEDURES.md, and QUICK_START.md are all cited but none exists — moving the Phase 6 tool catalog into TOOLS_REFERENCE.md would also slim the body.
Move the time-sensitive note "EWASCatalog ... returns HTTP 503, which it has since 2026-10-05" into a labeled status/deprecated section so the stale date does not read as permanent guidance.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Nearly all content is non-obvious operational knowledge rather than concepts Claude already knows, but the rows-vs-unique-sites pitfall is explained three times ("TOP-OF-MIND RULE", the PRIMARY SCRIPT field table, and the "Distinguish rows vs unique sites" table) with overlapping question-to-metric mappings, plus dataset-specific ranges ("typically 10k-30k") and an inline date-stamped note ("HTTP 503, which it has since 2026-10-05") that will go stale. Consolidating the three sections into one would remove significant duplication; it is above score 2 because there is no generic-concept padding. | 3 / 5 |
Actionability | The hot paths are fully concrete — a copy-paste bash invocation of the bundled `scripts/methylation_density.py` with a question-to-JSON-field mapping table, exact tool parameters (`gene_symbol` NOT Ensembl ID, case-sensitive `"Whole_Blood"`), and chi-square expected-count formulas. It falls short of 5 because Phases 1-5 are terse bullet descriptions with no executable code and defer to `CODE_REFERENCE.md`, which is absent from the bundle. | 4 / 5 |
Workflow Clarity | The Phase 0-7 sequence is explicit, RULE ZERO provides a clear precedence gate (check `*_executed.ipynb` and pre-computed results before re-analyzing), and the trickiest step has a feedback loop ("if your filtered count is two orders of magnitude smaller than the GT range ... re-run with the other axis and compare"). Not 5 because the other phases have no validation checkpoints; not 3 because sequencing and the key error-recovery loop are explicit. | 4 / 5 |
Progressive Disclosure | The body has clear section headers, point-of-use reference signals ("See ANALYSIS_PROCEDURES.md for decision tree"), and a real bundled script, but all four named reference files (`CODE_REFERENCE.md`, `TOOLS_REFERENCE.md`, `ANALYSIS_PROCEDURES.md`, `QUICK_START.md`) are missing from the bundle, and the ~30-line inline tool catalog in Phase 6 is content the intended TOOLS_REFERENCE.md was meant to hold. Structure is present and navigable (above anchor 2), but the dangling references keep it below anchor 4. | 3 / 5 |
Total | 14 / 20 Passed |