Content
71%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 router with lean prose, concrete rules, and a clearly sequenced workflow with an audit checkpoint. Its main defect is bundle integrity: the core layer files it instructs Claude to read first (manifest.yaml, static/core/*) are absent, so progressive disclosure fails at step 1 even though the references/ layer is complete.
Suggestions
Ship or verify the core layer files the body depends on — manifest.yaml, static/core/stance.md, static/core/workflow.md, and static/core/chinese-mode.md — or inline their essential content into SKILL.md, since the current bundle contains none of them.
Add an explicit feedback loop to the workflow (e.g., "if the FAIR audit finds gaps, fix the statement and re-run the audit before returning text") so the validation checkpoint functions as a recovery loop, not just a gate.
Trim the step-3 paragraph that re-enumerates the eight workflow steps already defined in core/workflow.md to remove duplicated guidance.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The router body is lean — no explanation of what DOIs or repositories are, and it delegates detail to references. Minor trimmable duplication: step 3 re-summarizes the eight-step workflow ("identify the journal, inventory every supporting dataset, classify each...") that is also stated to live in core/workflow.md. Not 5 due to that duplication; well above the verbose level-3 anchor. | 4 / 5 |
Actionability | Concrete instruction-only guidance: explicit load order ("Read [manifest.yaml](manifest.yaml). Then read every file listed under always_load"), an enumerated closed set of seven access routes, and hard rules ("Do not invent DOIs, accession numbers, repository names, licences..."). Not 5 because the actual drafting patterns and audit criteria live entirely in reference files, none of which are inlined, so a fresh run cannot draft anything until external files load. | 4 / 5 |
Workflow Clarity | A clear 4-step routing sequence plus an inline enumeration of the eight-step workflow ending in a validation checkpoint ("run the FAIR/metadata audit") and a defined return format ("ready-to-paste text plus unresolved fields"). Not 5: there is no explicit feedback loop (audit fails → fix → re-audit) and the full workflow details are deferred to a file that is not in the bundle. | 4 / 5 |
Progressive Disclosure | The design is good — on-demand references, one level deep, each named with its purpose (e.g., "references/policy-principles.md for the governing rules") and all seven references/ files named in the body exist. But judged against the actual bundle: the primary routing targets from step 1 and 2 (manifest.yaml, static/core/stance.md, static/core/workflow.md, static/core/chinese-mode.md) and ../nature-shared/journal-formats/nature-machine-intelligence.md do not exist here, so navigation breaks at the first instruction. Not 2 because the structure that is present is well-signaled and the references/ layer is real and correctly named. | 3 / 5 |
Total | 15 / 20 Passed |