Premortem-bind

A pre-mortem only matters if the surfaced failures bind to commitments. Naming "we could lose data" and moving on is decoration. This skill enforces binding: no risk leaves the room without an owner, a regression hook, and a renegotiation clause.

The 5 phases

1. Framing

State the assumption being stress-tested in one sentence. Format:

"This plan will succeed at <outcome> under <conditions>."

If you cannot phrase it in one sentence, the plan is not focused enough; go back to planning.

2. Free generation

Generate ≥8 failure modes. No filtering, no probability-weighing yet. Group as you go:

• Data / edge cases / encoding / time zones
• Security / auth / multi-tenant / privilege escalation
• Perf / scale / cold-start / N+1 / fan-out
• Ops / observability / rollback / on-call burden
• UX / accessibility / i18n / offline
• Concurrency / consistency / partial failure / retry storms
• Dependencies / supply chain / version skew

If you cannot list eight, you have not stressed enough. Spawn the grill-me skill or a fresh-context adversarial sub-agent to push further.

3. Behavioral attractor test

For each failure mode, ask:

"If I remove this top-1 failure and re-derive, does the same shape come back?"

• Yes → it is a structural attractor (architectural, will recur). Treat as a design constraint, not a discrete bug. May require re-planning, not just adding a test.
• No → it is a discrete risk to mitigate with a specific guard.

4. Probability × impact ranking

Sort all failures by P × I (each 1-5). Keep the top 5. Drop or defer the rest with a one-line reason each.

5. Commitment binding (mandatory — no skipping)

For each top-5 failure, produce three fields. If you cannot fill all three, the failure was decoration, not a risk — drop it or re-stress.

• Named owner — a concrete test file path, a sub-agent name, or a monitoring rule. Not "the team", not "ops". Examples:
  • tests/auth/session-fixation.test.ts
  • code-reviewer subagent on every auth-touching PR
  • Datadog monitor: payment_intent.failed rate > 1% over 5m
• Regression test hook — the concrete artifact (test file + assertion description) that fails when this failure mode occurs. If the test does not exist, write it now or schedule it as phase 1 of the plan.
• Renegotiation clause — what triggers revisiting this commitment?
  • "If P99 latency > 300ms for > 5m on the checkout path"
  • "If we add multi-region — the assumption changes"
  • "Never — this is a permanent constraint"

Output format (exact structure)

## Pre-mortem on: <one-sentence framing>

### Top failure modes (ranked)
1. <description>  P:<1-5>  I:<1-5>  type: <discrete | attractor>
2. ...
5. ...

### Bindings
| # | Failure                | Owner                       | Regression test               | Renegotiation                       |
|---|------------------------|-----------------------------|-------------------------------|-------------------------------------|
| 1 | Session fixation       | tests/auth/session.test.ts  | "old session ID invalidated"  | Permanent                           |
| 2 | ...                    | ...                         | ...                           | ...                                 |

### Re-plan signals (attractors)
- <attractor 1>: <what this implies for the plan — re-shape needed?>

### Deferred (with reason)
- <failure>: <why it's safe to defer>

Anti-patterns

• ❌ Generating <8 failure modes → you didn't stress enough.
• ❌ "The team will watch for this" — not an owner.
• ❌ A binding without a concrete test file path → the test doesn't exist, you haven't bound.
• ❌ Skipping the attractor test → you might patch the symptom of an architectural problem.
• ❌ Running a pre-mortem after coding → that's a retrospective, different purpose, different ritual.
• ❌ A pre-mortem with no top-1 failure (all "low P, low I") → either the plan is trivial (skip the skill) or you're not actually stressing.

Integration with other skills

• After slice-check accepts the plan → run premortem-bind on the highest-risk slice.
• If the pre-mortem surfaces an attractor → re-plan; do not patch.
• The bindings produced here feed directly into to-issues as test-writing tasks.

$ARGUMENTS