/premortem — Imagine the Incident

A fresh-context pass that assumes the latest work shipped and caused an incident, then works backward to the gap. Distinct from /agent-review: that audits what the work contains; this imagines what it is missing. The two are complementary lenses, not a repeated read.

Usage

• /premortem — review the latest work-unit commit (HEAD)
• /premortem <ref> — review a specific commit or range

Workflow

1. Load .claude/agents/premortem-agent.md.
2. Resolve the diff: default to the last commit (HEAD); use <ref> if given.
3. Dispatch named agent premortem-agent with the diff and one line of context on what the work unit did, and await its result before continuing. This skill must never return PASS before the review has actually run.
4. Surface the agent's verdict: PASS is silent; CONCERNS/BLOCK list each imagined incident, its mechanism in the diff, and the named missing guard, as follow-ups. Findings never revert the commit.

Dispatch context

When an orchestrator (/continue) dispatches this as a boundary review pass, the agent runs concurrently with agent-review — two independent fresh-context reads of the same diff. The timing and the diff it reads in that path (every commit of the just-completed scenario or task step, dispatched in that boundary unit's /refactor batch — it does not run on mid-block units) are owned by /continue; see its "Boundary Review Passes" section. A deterministic triage predicate runs before that dispatch and may skip both passes for an inert boundary. When they do run, /continue reads each finding's fixability tag and consumes the SAFE subset in an inline auto-fix (a separate review-fix: commit), while a NEEDS_CLARIFICATION finding — the last resort, gated by .claude/templates/workflow/clarification-escalation-test.md — is resolved by a single batched boundary quiz; see its "Triage & Auto-Fix" section. Run standalone, this skill reviews HEAD (or <ref>) on demand. The agent is read-only; it reports, it does not edit.