Content
70%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.
An exceptionally actionable, well-sequenced operational reference with explicit validation checkpoints and error-recovery loops. Its weaknesses are token efficiency (a load-bearing invariant repeated across four sections) and the absence of any progressive disclosure despite substantial specialized sub-topics that could live in reference files.
Suggestions
State the explicitly-configured-required-gate-unmet rule once (in Convergence policy) and reference it from Fail-open identity and Cross-references instead of restating it each time.
Move the delta-review mechanics, the closure-second-review contract block, and the audit-receipts detail into a references/ file, keeping SKILL.md to the core invocation, convergence, and verdict-folding flow.
Trim the second JSON example to only the fields that differ from the first (carried_legs, error_class, fail_open_backends) and show one filled claude_verdict object to make the invoke snippets fully copy-paste ready.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Dense with non-obvious domain rules, but the required-gate-unmet invariant is restated roughly four times (Convergence policy, Fail-open identity, Cross-references, Checking the result), and the second full JSON output example largely duplicates the first — more than the 'minor instances' of the 4 anchor. | 3 / 5 |
Actionability | Concrete throughout — exact tool name, full parameter table, copy-paste commands ('moai verify audit-plan --project-root <toplevel> --result-file <that path>') and the exact AUDIT-VERDICT line format. Minor gaps keep it below 5: the invoke snippets use unfilled placeholders (<your verdict>, <one-line>) with no worked claude_verdict example, and the digest members are listed in prose without a sample JSON object. | 4 / 5 |
Workflow Clarity | Clear sequencing via the path-selection decision table, an explicit 3-step validation procedure against the plan, and genuine feedback loops for error recovery (rerun_legs for inconclusive legs, FAIL → repair → delta re-audit chain) plus a 5-case verdict-folding table — matches the top anchor. | 5 / 5 |
Progressive Disclosure | No bundle files exist and all ~370 lines are inline; specialized branches a typical invocation does not need (delta review, the closure-second-review contract block, audit receipts) are inlined rather than split into reference files. Section headers are good, but content that should be separate is inline, matching anchor 3 rather than 4. | 3 / 5 |
Total | 15 / 20 Passed |