菜单
Consolidate findings from a sweep (UAT calibration, multi-agent validation, manual audit) into a TRIAGE.md document with severity ranking and ship recommendations. Use when a sweep has produced multiple findings that need to be triaged together to make a single ship/no-ship decision.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Stress-test a plan against the project's existing documentation, code, and decisions. Interview-style — one question at a time, recommend an answer, wait for feedback. Update PROJECT_BIBLE.md inline as terms and decisions are resolved. Offer ADRs only when the decision is hard to reverse. Use when a design needs sharpening before implementation.
Step back from the current task and produce a higher-level map of where the work fits — modules, callers, decisions, and where this sits in the project's phase. Use when unfamiliar with an area of code, between features, at a phase transition, or whenever the next decision needs context the immediate work doesn't supply.
Compact the current conversation into a session-boundary handoff document for another agent to pick up. Use at session end (or before a planned interruption like a machine reboot, hard context-window break, or hand-off to a different operator).
| name | sweep-triage |
| description | Consolidate findings from a sweep (UAT calibration, multi-agent validation, manual audit) into a TRIAGE.md document with severity ranking and ship recommendations. Use when a sweep has produced multiple findings that need to be triaged together to make a single ship/no-ship decision. |
| argument-hint | Path to the sweep directory (e.g., Reports/uat-2026-MM-DD-foo/) OR a description of the sweep being triaged. |
What this is. A skill for consolidating findings from a sweep into one decision-grade document. A sweep is any activity that surfaces multiple findings at once: a multi-agent UAT run, a manual audit pass, a security review, a calibration replay. Each finding needs severity, a fix recommendation, and the whole set needs an overall ship recommendation.
What this is NOT. This skill is not for per-issue triage in an issue tracker (working through GitHub Issues or Linear one at a time, applying state-machine labels). That is a different workflow. If you want issue-tracker triage, see the related work below.
Canonical example.
Reports/uat-2026-04-29-bl029-validation/TRIAGE.md(committed to this repo) is the output shape this skill produces. Read it before writing your first TRIAGE — it shows the per-agent verdict table, severity-ranked issue list, and the Resolution section appended after fixes shipped.
Reports/<sweep-id>/results/agent-*.json.The skill produces a markdown document at:
Reports/<sweep-id>/TRIAGE.md
Where <sweep-id> is the directory the sweep already lives in (e.g., uat-2026-04-29-bl029-validation). If there is no sweep directory yet, create one at Reports/<sweep-id>/ first.
The TRIAGE.md MUST contain, in this order:
ship | fix-then-ship | block-merge), headline finding.S1 (critical / merge-blocker) through S5 (low / nice-to-have). Each issue gets:
fix-in-branch | defer-to-followup | accept-as-is | block-mergeLocate or create the sweep directory. If the user passed Reports/<sweep-id>/, use it. Otherwise infer from context (current sweep in flight) or ask.
Gather findings. Inspect everything in the sweep directory that produces signal:
results/agent-*.json) — each typically has a ship_recommendation field and a list of findings.Cluster duplicate findings. When multiple agents report the same underlying issue with different framings, merge them into one issue with all sources cited. Don't list the same problem 3 times.
Assign severity. Use the ladder:
refuse_to_recommend audit rows (BL-029 calibration S1); secret leakage; data loss path.Recommend per-issue action. For each issue, pick: fix-in-branch (do it now), defer-to-followup (file a BL item), accept-as-is (document why and move on), block-merge (don't ship until this is fixed).
Recommend overall ship decision. Bias toward shipping with a clear deferred list over blocking everything. Block-merge should be rare — only when an S1 is unresolved or a cluster of S2s collectively crosses the credibility threshold.
Write "What demonstrably works." This is the section that prevents misreading. Examples: "Confirmation phrase defeats novice acceptance (agent 4 confirmed in-character)" or "Severity elevation routes through correctly when patterns DO match."
Save the TRIAGE.md. Commit it as part of the sweep evidence, not as a transient artifact.
After fixes ship — append Resolution. Once the recommended fixes are in, return to the TRIAGE and append a Resolution section recording exactly what was done (commit refs) and what was deferred (BL items). This converts the TRIAGE from a snapshot into a permanent record.
Read the document with fresh eyes. Ask:
fix-in-branch this subset, defer-to-followup as BL-029.1) or vague (investigate)?If any answer is no, fix it before saving.
solo-orchestrator-backlog.md): defer-to-followup items should be filed as BL items. Cite the BL number in the TRIAGE so the reader can trace.docs/superpowers/plans/): if a fix-in-branch cluster is large enough to warrant a plan, write one. The TRIAGE points to the plan; the plan points back to the TRIAGE.Reports/<sweep-id>/): the TRIAGE lives in the same directory as the agent reports it triages.scripts/escalate-to-user.sh (BL-029) to write a structured pending-approval rather than burying the question in the TRIAGE.mattpocock/skills triage (upstream, MIT, © Matt Pocock) — a spiritually-related but differently-scoped skill for per-issue triage in an issue tracker (GitHub Issues, Linear). State machine model with needs-triage / needs-info / ready-for-agent labels. Use that if you're working through a backlog of incoming issues one at a time; use this skill if you're consolidating a sweep into a single ship decision.session-handoff (Solo, adapted from mattpocock) — what to do when the work is mid-flight and the session is ending. Different artifact; different purpose.