Plan review system with three adversarial subagents (Advocate, Analyst, Critic) that debate a proposed plan from different perspectives. Use when reviewing implementation plans, architecture decisions, or significant changes. Triggers: triumvirate, debate plan, review plan, argue plan, adversarial review, devil's advocate, plan critique, three perspectives, plan debate.
<command-name>triumvirate</command-name>
Three adversarial subagents debate a proposed plan, then the plan is amended with the strongest arguments. The user decides how to proceed.
| Reviewer | Role | Perspective |
|---|---|---|
| Advocate (+) | Argue FOR the plan | Strengths, opportunities, enhancements, counter-arguments to criticism |
| Analyst (=) | Objective tradeoff analysis | Assumptions, alternatives comparison, metrics, risk/reward |
| Critic (-) | Find weaknesses constructively | Risks, edge cases, tech debt, failure modes — every criticism gets a mitigation |
User Plan → [Advocate + Analyst + Critic in parallel] → Synthesis → Amended Plan → User Decision
Extract the plan from current conversation context:
Launch all three with the Task tool simultaneously:
Task(subagent_type='general-purpose', prompt=advocate_prompt)
Task(subagent_type='general-purpose', prompt=analyst_prompt)
Task(subagent_type='general-purpose', prompt=critic_prompt)
For the full prompt for each reviewer, see:
Replace {plan_text} in each prompt with the actual plan content.
After all three return, synthesize arguments and amend the plan. For templates, see templates.md.
Present findings directly — do NOT use AskUserQuestion. Wait for user response.
Options presented after review:
/triumvirate again on the amended planEach subagent MUST research before forming arguments:
| Type | Minimum | Tools |
|---|---|---|
| Codebase research | 3 searches | Glob, Grep, Read |
| Web research | 2 searches minimum, always required | WebSearch, WebFetch |
| File reads | 2 files | Read |
Web research is not optional. Each reviewer must back its arguments with research into standard industry practices for similar systems or features. Prefer strong engineering sources when available:
If exact FAANG-style analogues do not exist, use the closest high-quality engineering sources and say so explicitly.
/triumvirate → three subagents debate