Content
81%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, disciplined instruction skill: an explicit gated workflow with strong fail-fast validation, concrete templates and worked examples, and condition-based reference pointers that keep the body lean. The main costs are duplicated maxims across sections, steps that lean on external contracts without inline summaries, and a reference list with repeated and out-of-bundle paths.
Suggestions
Remove the verbatim duplication of 'Linear is execution state; .harness is cognition and proof' (keep it in Philosophy, drop the Gotchas restatement) and merge the overlapping untrusted-text rules in Constraints and Safety Boundaries into one section.
Consolidate the References list so each file appears once with its combined read conditions, removing the repeated entries for linear-filing-rule.md and source-prompt-preservation.md.
Inline one-line summaries of the externally-defined policies invoked by procedure steps (e.g. 'shared subagent policy', 'XP value filters') or point to the specific reference section that defines them, so the steps are executable without a blind file load.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The telegraphic style assumes Claude's competence and adds domain-specific rules rather than general knowledge, but there is real redundancy: 'Linear is execution state; .harness is cognition and proof' appears verbatim in both Philosophy and Gotchas, the untrusted-text rule is duplicated across Constraints and Safety Boundaries, and the mutation-confirmation gate is restated in four sections. | 4 / 5 |
Actionability | Concrete guidance throughout — named templates (Bug, Feature, Research, Release, Governance/Policy), an exhaustive list of required output fields, an executable validation command, three worked examples with real artifact paths, and exact refusal wording — but several procedure steps defer to undefined external contracts ('apply shared subagent policy', 'XP value filters') rather than stating the guidance inline. | 4 / 5 |
Workflow Clarity | The 12-step procedure is clearly sequenced from classification through live-state verification, confirmation gating, drafting, and mutation, with explicit fail-fast validation ('stop at the first failed gate... until the failure is fixed, waived... or reported as blocked'), blocked-status error handling, and a dedicated Failure Mode section — a full feedback loop for a mutation-capable skill. | 5 / 5 |
Progressive Disclosure | The body stays an overview and the References section uses well-signaled, condition-based 'Read when X' pointers, with the three local bundle files (references/contract.yaml, references/evals.yaml, references/task-profile.json) real and correctly referenced. However, ~14 pointers include duplicate entries for the same files and many paths escape the skill directory via '../../references/**', which weakens navigation and bundle self-containment. | 4 / 5 |
Total | 17 / 20 Passed |