Content
85%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A dense, well-structured instruction skill with a concrete output contract, a sequenced workflow with an explicit QA checkpoint, and clean one-level-deep reference navigation. The only weakness is redundancy between the stance and red-lines sections.
Suggestions
Consolidate the duplicated prohibitions between "Default stance" and "Red lines" (reviewer-identity invention and editorial-decision / Nature-fit claims) into a single canonical list to tighten token use.
Consider trimming the "Accepted inputs" enumeration, which lists mostly inferable input types, or move it into a reference file.
Add a brief "if QA fails, fix and re-run" feedback note to step 9 to make the validation loop explicit rather than implied.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly lean and free of concept-padding, but "Default stance" and "Red lines" duplicate several prohibitions (inventing reviewer identities, claiming editorial decision / Nature fit) that could be consolidated; not quite "every token earns its place". | 2 / 3 |
Actionability | Provides a concrete 9-step workflow and a copy-paste-ready output template with exact fields (Concern ID: R1-M1, Axis, Claim pointer, Evidence pointer), plus specific rules such as consensus requiring two reviewers. | 3 / 3 |
Workflow Clarity | A clearly numbered 9-step sequence culminates in an explicit QA validation step ("Run QA for evidence anchoring, overlap, groundedness...") supported by the qa-checklist reference, above the implicit-checkpoint bar. | 3 / 3 |
Progressive Disclosure | Overview body points to 9 verified one-level-deep reference files via a well-signaled "Related files" table with "Open when" conditions and a source hierarchy; all referenced paths exist and navigation is easy. | 3 / 3 |
Total | 11 / 12 Passed |