Content
73%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.
The body delivers a rigorously sequenced, highly actionable read-only comparison workflow with explicit validation checkpoints and a clear verdict contract, supported by real, well-signaled one-level references. Its main cost is token efficiency: repetitive guardrail restatements and dense compound sentences make it harder to parse than the underlying rules require.
Suggestions
Consolidate the repeated no-read/no-deploy/no-collect guardrails (currently restated in the availability gate, intent gate, section 2, and section 5) into a single scope-constraints statement referenced from each section.
Break the long compound-negation sentences (e.g. 'A yes/"continue" permits comparison of existing explicitly supplied evidence only, not new reads, jobs, traffic or saving') into short bullet constraints to improve parseability at the same token cost.
Use a platform-neutral path in the invocation (e.g. 'python "<skill-directory>/scripts/recovery.py"') so the command is executable on POSIX systems as well as Windows.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body never explains concepts Claude already knows and nearly every sentence carries a rule, but it is noticeably padded with repetition: the no-read/no-deploy/no-collect guardrails are restated across the availability gate, intent gate, section 2, and section 5, and long compound-negation sentences (e.g. 'A yes/"continue" permits comparison of existing explicitly supplied evidence only, not new reads, jobs, traffic or saving') could be tightened. It fits anchor 3 (mostly efficient but could be tightened) better than anchor 2, since the excess is redundancy rather than irrelevant explanation. | 3 / 5 |
Actionability | Concrete, executable guidance throughout: an exact invocation ('python "<skill-directory>\scripts\recovery.py" <sanitized-input.json>'), explicit arithmetic rules ('compare total / sample_count against the explicitly agreed bound', 'compare the explicitly supplied source-computed percentile_value directly'), and a definitive verdict enum (VERIFIED_RECOVERED / NOT_RECOVERED / INCONCLUSIVE). Minor gaps keep it from anchor 5: the path uses a Windows-only backslash separator, and no worked input example is shown inline (deferred to contract.md and fixtures). | 4 / 5 |
Workflow Clarity | A clear sequence (availability gate, intent gate, entry contract, then numbered steps 1-5 from scope agreement through reporting to optional follow-up) with explicit validation checkpoints: thresholds agreed with the task owner 'before execution', original-window/post-window chronology rules, INCONCLUSIVE on missing or ambiguous evidence, and an error-recovery loop ('Invalid helper input instead produces a sanitized error/nonzero exit; correct it explicitly, never call it recovered'). Verdict precedence rules are explicit, matching anchor 5. | 5 / 5 |
Progressive Disclosure | Direct references are real files, one level deep, and each is clearly signaled with its purpose ('Read the evidence contract before constructing input'; fixtures marked as 'synthetic, not recommended SLOs'; the helper script's constraints stated inline). Minor gaps: a large amount of dense policy is inlined in SKILL.md that could live in a reference, and the shared/ tree (next-actions.md chaining to handoff.md, trace-capabilities.md, and a nested investigation/ subtree with its own references/ and scripts/) creates 2+ level chains, consistent with anchor 4 rather than 5. | 4 / 5 |
Total | 16 / 20 Passed |