Content
56%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 has a genuine skeleton of a good skill — sequenced procedure, explicit validation gates, fail-fast failure handling, and one concrete executable command — but execution guidance stays abstract, navigation is cluttered by an eight-times-repeated pointer line and a prose-heavy References section, and six ../../ references dangle unresolved. It reads as a compressed internal contract spec rather than an operational guide.
Suggestions
Replace the eight repeated 'See references/hot-path-folded-context.md for folded X detail' lines with a single pointer at the top or bottom of the skill, and consolidate the duplicated 'memory, artifact, goal, Chronicle, Local Memory, vault, or native citation surfaces' enumerations into one defined list.
Add one worked example of a captured learning artifact (frontmatter, solved evidence, validation output, prevention note) — or point to assets/resolution-template.md from the body — so the capture/refresh procedure is executable rather than a list of eligibility rules.
Fix the six dangling '../../references/*' paths by copying those files into this skill's references/ or correcting the relative paths, and restructure the References section as a bulleted list of 'read when -> file' pairs.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is compressed rather than padded — it never explains concepts Claude already knows — but it carries real waste: the boilerplate line "See references/hot-path-folded-context.md for folded X detail" is repeated in eight sections, and the enumeration "memory, artifact, goal, Chronicle, Local Memory, vault, or native citation surfaces" is repeated almost verbatim in two places. This is 'mostly efficient but could be tightened' rather than anchor 2, since the padding is duplicated pointers, not unnecessary explanation. | 3 / 5 |
Actionability | Concrete elements exist — the exact command "python3 Plugins/harness-engineering/scripts/check_bluf_structure.py <learning-artifact-path> --json" and a fully enumerated structured-output field list — but the core procedure steps are abstract directives ("Prove eligibility", "Keep scope tight", "Avoid duplicate memory") with no worked example, template, or artifact sample. This matches anchor 3 (some concrete guidance but incomplete) better than 4, whose 'concrete code or commands with minor gaps' would require the primary workflow itself to be executable, not just its validation script. | 3 / 5 |
Workflow Clarity | The Procedure is a numbered five-step sequence, and the Validation section adds explicit fail-fast gates with per-gate pass/fail/blocked reporting plus a blocked_reason/next_evidence_needed recovery path in Failure Handling — checkpoints are present, not implicit. It falls short of anchor 5 because most gates name what to check (solved evidence, redaction, discoverability) without any method or re-validate loop beyond the single structure script; it is above anchor 3 because validation and error handling are explicit sections. | 4 / 5 |
Progressive Disclosure | Local references are real and well-signaled with read-when conditions (contract.yaml, evals.yaml, source-prompt-preservation.md, compound-learning-migration.md all exist), but the References section is a tangle of prose-like bullets, and six cited paths point outside the bundle to "../../references/*" (codex-native-memory-baseline.md, stage-arc-boundary-contract.md, bluf-review-contract.md, subagent-call-contract.md, visual-reference-contract.md, deferred-context-index.md) that do not resolve from the skill directory. That mix of working but disorganized navigation and dangling paths fits anchor 3 (some structure, could be better organized) rather than 2, since references are explicitly signaled rather than buried. | 3 / 5 |
Total | 13 / 20 Passed |