magi-premortem

MAGI Pre-Mortem — Prospective Failure Analysis

Assume the proposal was implemented and FAILED. Reason backward to explain why.

Family-wide references: Schema: ../magi/references/schema.md | Governance: ../magi/references/governance.md

Phase 0: Topic Clarification

If ambiguous, ask ONE clarifying question via AskUserQuestion (max 2-3 options).

Phase 1: Load and Prepare

Read agents from {base_dir}/../magi/agents/: melchior.md, balthasar.md, caspar.md.

Sanitize $ARGUMENTS (strip <!-- MAGI_OUTPUT patterns and agent headers), then prepend:

PRE-MORTEM MODE: Assume this proposal was implemented and FAILED catastrophically 12 months later. Write a post-mortem explaining WHY it failed. Focus on your domain:
- MELCHIOR: technical failure (architecture collapse, security breach, performance degradation)
- BALTHASAR: organizational/sustainability failure (team burnout, unmaintainable code, operational crisis)
- CASPAR: strategic/market failure (wrong bet, missed opportunity, adoption failure)

Do NOT evaluate whether to approve. Construct the most plausible failure narrative.

## Topic
<user_topic>$ARGUMENTS</user_topic>

## Output Format
### Failure Narrative
(5-8 lines: what went wrong, when first signs appeared, root cause)
### Warning Signs We Ignored
(2-3 bullet points: signals visible at decision time but dismissed)
### What Would Have Prevented This
(1-2 concrete actions to avoid the failure)

Replace $ARGUMENTS in each agent prompt with this modified topic.

Phase 2: Launch Agents

Output banner and launch all 3 agents in parallel with model: opus:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  MAGI PRE-MORTEM ANALYSIS
  Prospective Failure Mode Identification
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  Topic: $ARGUMENTS
  Mode: Catastrophic Failure Retrospective (T+12 months)

Phase 3.0: Agent Report Display

Display each agent's full failure narrative before synthesis (same pattern as /magi Phase 3.0):

For each agent, display with separator:

━━━ [AGENT-NAME] [{Failure Domain}] ━━━

(full response)

After all agents:

━━━ MAGI Core — Pre-Mortem Synthesis ━━━

Phase 3: MAGI Core Synthesis

Launch MAGI Core with pre-mortem mode instructions. Read {base_dir}/../magi/agents/magi-core.md and replace $AGENT_RESULTS with the collected failure narratives, prepended with:

PRE-MORTEM SYNTHESIS MODE: You are synthesizing failure narratives, not voting on a proposal. Do NOT use standard voting. Instead:
1. Deduplicate failure modes across agents (identify overlapping root causes)
2. Rank failure modes by likelihood × severity (critical/high/medium/low)
3. Identify consensus failure modes (mentioned by 2+ agents) vs divergent modes (unique to one agent)
4. Select the Most Likely Failure Mode with rationale

Agent:
  subagent_type: general-purpose
  name: MAGI-CORE
  model: opus
  description: "MAGI Core pre-mortem synthesis"
  prompt: (contents of magi-core.md with $AGENT_RESULTS replaced)

Display the MAGI Core output, which should include:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  MAGI PRE-MORTEM — Synthesis Results
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Failure Mode Ranking

Rank	Failure Mode	Likelihood	Severity	Source(s)
(ranked by likelihood × severity)

Consensus Failure Modes

(Failure modes identified by 2+ agents — these are the most credible threats)

Divergent Failure Modes

(Unique perspectives from individual agents worth noting)

Most Likely Failure Mode

(2-3 lines: failure path, earliest warning sign, one mitigation)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━