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 well-structured runbook body: a clearly sequenced 14-step workflow with explicit validation checkpoints and a failure-recovery loop, concrete commands and decision tokens, and conditionally routed one-level-deep references. Weaknesses are minor: some repetition of contract-preservation and validation rules across sections, output formats deferred entirely to references, and three bundle files (contract.yaml, evals.yaml, task-profile.json) left unrouted by anything more specific than a catch-all line.
Suggestions
Replace the vague catch-all reference line with one routed entry per bundle file (e.g., 'Read references/contract.yaml when X', 'Read references/evals.yaml when Y') so all six files in references/ are discoverable by condition, closing the progressive-disclosure gap.
Consolidate the contract-preservation rule (currently stated in Workflow step 9, Failure Mode, and the refusal bullets) into a single authoritative location cross-referenced from the others, trimming the conciseness overlap.
Show a minimal inline example of a Context ledger entry and a verified Context Pointer alongside the Outputs section, or point to the exact reference section that defines their formats, to make the output guidance copy-paste actionable.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with skill-specific policy and explains nothing Claude already knows, fitting the efficient 4 anchor; however, contract-preservation and validation rules are repeated across Workflow steps 9 and 14, Failure Mode, and Validation, which could be consolidated, keeping it below the lean 5 anchor. | 4 / 5 |
Actionability | Concrete executable guidance is present — exact commands ('rg --files', 'rg -n "(command|path|handle)" .'), explicit decision tokens ('Decision required:', 'no_justified_edit', 'schema_version'), and named outputs — with minor gaps such as the Context ledger and Context Pointer formats being routed to references rather than shown. This matches the mostly-executable 4 anchor, not the copy-paste-complete 5 anchor. | 4 / 5 |
Workflow Clarity | The 14-step sequence is clearly ordered with an explicit validation step (14), a dedicated Validation section with pass/fail/blocked reporting, and an explicit feedback loop ('If validation fails, fix the first failure class and rerun the same gate before broadening'), matching the 5 anchor including error recovery for a refactor-style skill. | 5 / 5 |
Progressive Disclosure | The three routed references exist, are one level deep (verified — none chain to further reference files), and each is conditionally signaled ('Read references/discovery-interview.md only when the request is underspecified'). However, references/ also contains contract.yaml, evals.yaml, and task-profile.json, which are covered only by the vague catch-all 'Load other archived references... only when the active workflow requires that exact detail' — a minor navigation gap placing this at the 4 anchor rather than 5. | 4 / 5 |
Total | 17 / 20 Passed |