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.
Strong domain-specific operational guidance with a clearly sequenced, checkpoint-bearing 8-phase workflow and genuinely useful interpretation frameworks (evidence grading, SVG vs. cell-type-marker reasoning). Weaknesses are self-duplication (workflow diagram vs. phase summaries, platform facts vs. limitations), inlined peripheral dataset-discovery material, and references to bundle files that are not present.
Suggestions
Collapse the ASCII "Workflow Overview" and the "Phase Summaries" into a single presentation — keep the diagram and fold each summary's extra specifics (PCA components, Wilcoxon, method recommendations) into it, cutting roughly 40 duplicated lines.
Move the HuBMAP biospecimen/donor discovery layer (~70 lines, peripheral to spatial analysis) into a dedicated reference file and keep only the availability note plus one fallback example inline in SKILL.md.
Ship code_examples.md and report_template.md in the bundle — or inline minimal executable snippets for Phases 1-6 — so the core analysis code the body defers to those files is actually reachable.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly dense, non-obvious operational detail (QC cutoffs "min 200 genes, 500 UMI, <20% MT", tool params, interpretation thresholds), but tightening is needed: the 8-phase workflow is rendered twice (the ASCII "Workflow Overview" diagram plus the "Phase Summaries" section), the opening paragraph restates the frontmatter description, "Limitations" repeats platform facts already in "Supported Platforms" ("Visium spots contain multiple cells"), and the HuBMAP section carries a second redundant fallback code block. Not 2 because the bulk is genuinely operational knowledge rather than padding; not 4 because the duplication is more than minor. | 3 / 5 |
Actionability | Mostly executable guidance: concrete QC thresholds, named tools with parameters ("OmniPath_get_ligand_receptor_interactions — ... Use `partners` param for specific genes"), runnable Python snippets (scanpy loading, HuBMAP queries), and interpretation cutoffs ("Moran's I > 0.3 with FDR < 0.05", "z-scores > 2"). Not 5 because the core Phases 1-6 have no inline code — it is deferred to code_examples.md — so the end-to-end common case is not copy-paste ready in the body; not 3 because everything present is real, specific, and executable rather than pseudocode. | 4 / 5 |
Workflow Clarity | A clear 8-phase sequence with most checkpoints explicit: QC filters and "Verify spatial alignment" (Phase 1), "Filter significant spatial genes (FDR < 0.05)" (Phase 4), "Validate with marker genes" (Phase 6), a fallback loop for unavailable HuBMAP tools, and a "Quantified Minimums" completion checklist. Not 5 because there is no error-recovery guidance (e.g., what to do when fewer than 500 spots survive QC or when deconvolution fails); not 3 because validation checkpoints are stated explicitly rather than implied. | 4 / 5 |
Progressive Disclosure | The body is well-sectioned with clearly signaled one-level references ("See [report_template.md](report_template.md)" and the "Reference Files" section), but the two referenced bundle files (code_examples.md, report_template.md) are absent from the bundle, leaving the deferred core-phase code unreachable; additionally ~70 lines of peripheral HuBMAP sample/donor discovery material are inlined in SKILL.md where a separate reference file would fit. Not 4 because dangling references plus inlined peripheral content exceed 'minor organization gaps'; not 2 because the document is clearly structured with headers, not a monolithic wall. | 3 / 5 |
Total | 14 / 20 Passed |