Content
77%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 an exceptionally actionable, tightly sequenced workflow with real validation and feedback loops — among the strongest execution protocols a multi-step skill can have. Its weaknesses are redundancy (repeated rules and emphatic padding) and the absence of any progressive disclosure: everything, including a reusable review-spec payload, lives inline in one long file.
Suggestions
Move the review dimensions (A–K), GTFOBins procedure, and findings output format into a references/review-instructions.md file and have the body instruct agents to read and forward it, shrinking SKILL.md to the orchestration loop.
Delete the closing "Important rules" section or reduce it to only rules not already stated in-step, and tighten the STOP section's repeated "Do NOT skip" bullets into a single sentence.
Trim meta-commentary like "The randomized order ensures that each run... avoiding systematic bias" — the rationale can be one clause instead of a paragraph.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The A–K review dimensions encode genuinely project-specific rules (callCtx.OpenFile, the 1MB output limit, skip_assert_against_bash semantics) that earn their tokens, but the closing "Important rules" section restates rules already stated in-step ("Sandbox first", "Stream, don't buffer", "Respect the batch limit", "Use gate checks") and the STOP section's repetitive "Do NOT skip... because you think..." bullets are padded emphasis. Not 4 because these redundant spans are more than minor trimmable instances. | 3 / 5 |
Actionability | Fully copy-paste-ready commands throughout: `gh pr view $ARGUMENTS --json number,url,headRefName,baseRefName`, `go test ./tests/ -run "TestShellScenarios" -timeout 120s`, exact commit-message formats (`[<target>] Fix null byte handling...`), exact PR-comment templates, and an exact findings output format with priority levels. Not below 5 because the concrete examples cover the common cases end to end. | 5 / 5 |
Workflow Clarity | Clear gated sequence (Step 1 → 2 loop → 3 → 4) with explicit TaskList verification before each step, feedback loops for error recovery (test failure → fix → re-run with a 3-attempt cap; per-target fix loop; full-sweep re-review in Step 3), a decision table for continuation, and a hard 50-batch limit. This is a batch operation with complete validation, matching the top anchor. | 5 / 5 |
Progressive Disclosure | Section headers and internal navigation are good, but the skill is a single monolithic file with no bundle references at all, and the ~100-line agent review-instruction spec (dimensions A–K, GTFOBins procedure, output format) — which is pasted verbatim into every subagent prompt — clearly belongs in a one-level-deep references file. Not 2 because the structure is genuinely strong, not minimal; not 4 because content that should be split out is fully inlined. | 3 / 5 |
Total | 16 / 20 Passed |