Content
63%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 is a lean, well-signaled procedural contract with explicit validation and fail-fast gates, but it is guidance-only with no commands, code, or worked output examples, and its progressive disclosure breaks in practice because every referenced detail file lies outside the bundle under three inconsistent path schemes. Actionability and reference resolution are the main gaps.
Suggestions
Add one worked example of structured output (a severity-ranked finding with file:line, traceability, verdict, and safe_to_continue fields) so the required output shape is concrete rather than only enumerated by field name.
Fix reference paths so they resolve within the skill bundle (e.g., move review-mode-contract.md, review-policy-index.md, and review-loop-patterns.md into the skill's references/ directory and reference them relatively), and unify the three inconsistent path schemes (../../references/, Plugins/harness-engineering/references/, Infrastructure/references/).
Consolidate the repeated "green CI is not readiness" guidance (Procedure step 7, Gotchas, and Anti-Patterns) into a single location to recover tokens.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and assumes Claude's competence — it never explains what a PR, diff, or CI is, and every section is directive. Not anchor 5 because the "green CI is not readiness" point is repeated across Procedure step 7, Gotchas ("Green CI is not readiness..."), and Anti-Patterns ("Treating green CI as proof of real user behavior"), which could be consolidated. Not anchor 3 because there is no conceptual padding or unnecessary explanation. | 4 / 5 |
Actionability | Concrete elements exist — the enumerated ten-mode list, required output fields (severity-ranked `file:line` findings, `blocked_reason`, `safe_to_continue`, `traceability`), and a numbered procedure — but there are no commands, no executable code, and no worked example of a finding or verdict; the Examples section shows input prompts rather than output shapes. Anchor 4 requires concrete code/commands with only minor gaps, which this lacks, while anchor 2's "minimal concrete guidance" understates the specific mode/evidence/output rules present. | 3 / 5 |
Workflow Clarity | A coherent 13-step sequence (mode selection → stage-context resolution → untrusted-evidence handling → evidence reading → findings → contracts → verdict) with explicit checkpoints: "Fail fast: stop at the first failed gate and do not proceed" and a Failure mode section with recovery ("stop and return the blocker with the smallest recovery step"). Not anchor 5 because steps 9–12 are conditional branches with ambiguous ordering and the validation section is two sentences rather than a gate-by-gate checklist; not anchor 3 because validation and recovery checkpoints are explicit, not implicit. | 4 / 5 |
Progressive Disclosure | The References section signals each external file with an explicit when-to-read condition, which is well-organized signaling. However, scored against the actual bundle, none of the referenced paths resolve: `../../references/skills/he-code-review/review-mode-contract.md`, `../../references/stage-arc-boundary-contract.md`, and all `Plugins/harness-engineering/references/*.md` and `Infrastructure/references/...` paths are absent from the bundle (which contains only contract.yaml, evals.yaml, task-profile.json), and three inconsistent path schemes are mixed. Not anchor 4 because broken, inconsistent reference paths are more than minor organization gaps; not anchor 2 because detail is properly deferred to external files rather than inlined and each reference is clearly signaled. | 3 / 5 |
Total | 14 / 20 Passed |