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.
A strong, highly actionable triage skill: executable commands and signal-to-action tables throughout, a well-checkpointed workflow with validation and fallback paths, and approval-gated safety rules for destructive operations. The only weaknesses are mild redundancy around outage checking and conservative use of its own references directory.
Suggestions
Consolidate the platform-outage guidance: state the two status-page URLs once (e.g., in 'Rule out a platform outage first') and have the classification table and base-rate section reference that rule instead of repeating it.
Add a short 'Running unattended' pointer in the SKILL.md body (e.g., under a '## Unattended master-red diagnosis' heading) to references/master-red-incident.md, so the body — not just the frontmatter description — signals the bundle's reference files.
Consider moving the merge-queue branch mechanics and base-rate query details into a reference file, keeping the digest-state table and classification table inline as the core triage surface.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Efficient and dense with non-derivable operational knowledge — merge-queue branch mechanics ('trunk-merge/pr-<n>/<uuid>' branches), digest state semantics, Depot log-404 behavior — with no padding about concepts Claude already knows. Not 5: the platform-outage guidance repeats three times ('Rule out a platform outage first', the infra/runner classification row, and the base-rate section's 'Check <https://www.githubstatus.com/>'), and a few conversational asides ('It costs two page loads', 'GitHub Actions goes down often enough that it belongs before any log reading') could be trimmed. Not 3: nearly every sentence carries repo-specific fact rather than unnecessary explanation. | 4 / 5 |
Actionability | Fully executable throughout: copy-paste commands ('hogli ci:insights search "<error or test name>"', 'gh run view <run-id> --log-failed', 'trunk merge status <n>', 'gh pr view --json number,headRefName,statusCheckRollup'); a signal-to-class-to-first-action table ('AssertionError, test diff… → reproduce with hogli test <path>::<test>'); a repro table with exact commands per class; and a copy-ready report template with explicit line limits ('Keep excerpts under 40 lines'). Specific examples cover the common cases. | 5 / 5 |
Workflow Clarity | Clear sequence with explicit validation checkpoints and feedback loops: outage check first ('a platform incident makes every other signal a symptom'), then insights, target resolution ('If neither works, ask the user for a PR URL or run ID. Do not guess.'), classification, base rate, repro, report. Validation is explicit ('confirm from run data before reporting a failure as already fixed', 'Check the current queue run with gh before reporting it as active', 'do not assert [transient] from one run'), with error-recovery fallbacks (exit 78 → gh-based inspection; '--log-failed' insufficient → full job log; MCP bounded by Logs retention → gh). Destructive actions are gated behind an explicit-approval safety checklist, and the report shape is a checklist. | 5 / 5 |
Progressive Disclosure | Good structure: well-organized sections, and the bundle's two reference files are real, clearly scoped, and one level deep (master-red-incident.md points to master-red-workflow-setup.md and back to the parent SKILL.md without chaining content). Not 5: the body itself never links its own references/master-red-incident.md — the pointer lives only in the frontmatter description — and master-red-workflow-setup.md is discoverable only through incident.md; 300+ lines of deep detail (merge-queue mechanics, base-rate analysis) also stay fully inline. Not 3: navigation is easy, references are real and well-signaled from the description, and content is not buried. | 4 / 5 |
Total | 18 / 20 Passed |