Content
88%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 exemplary operational runbook: concrete commands, an exact data-entry contract, sequenced steps with real validation, and honest anti-goals. The only weaknesses are transient state inlined into Hard rules (a staleness risk) and a body long enough that some reference material could be split out.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean and imperative with zero explanations of concepts Claude already knows, but Hard rule 3 bakes time-sensitive state ("the crush binary is now on PATH", the quoted standing next_action) directly into the body where it will go stale rather than placing it in a standing-targets reference. | 4 / 5 |
Actionability | Fully executable throughout: copy-paste commands (the jq cell query, HOME="$PWD" just harness-probe/--generate/--check), an exact field contract (parity, mode, surface, note, evidence=Evidence(kind=..., ref=..., version=..., date=...)), enumerated probe kinds, and an explicit whitespace-split argument grammar with a stop-and-report fallback. | 5 / 5 |
Workflow Clarity | Six clearly sequenced steps with p0/p1/p2 priority ordering and an explicit validation feedback loop: step 6 runs harness-check, which runs validate_registry() and names the offending <capability>[<slug>] on failure, and the "What not to do" section guards against integrity violations. | 5 / 5 |
Progressive Disclosure | Sections are well organized with no buried references, and the one external pointer (.agents/skills/optimize-harness/references/session-usage-evidence.md as the permalink pattern) is one level deep and clearly signaled. However, at ~90 dense lines the body inlines material (the crush standing target, the argument contract) that could live in reference files rather than the main skill. | 4 / 5 |
Total | 18 / 20 Passed |