Triumvirate

User Plan → [Advocate + Analyst + Critic in parallel] → Synthesis → Amended Plan → User Decision

Implementation

Step 1: Parse the Plan

Extract the plan from current conversation context:

• Search for plan files, implementation proposals, architecture decisions
• If no plan found, review the most recent significant discussion topic
• Do NOT prompt the user — always work with current context

Step 2: Launch Subagents in Parallel

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:

• Advocate Prompt
• Analyst Prompt
• Critic Prompt

Replace {plan_text} in each prompt with the actual plan content.

Step 3: Synthesize and Amend

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.

Step 4: User Decision

Options presented after review:

• Re-debate: /triumvirate again on the amended plan
• Approve: Begin implementation
• Modify: User edits manually
• Reject: Start over

Research Requirements

Each subagent MUST research before forming arguments:

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:

• large-scale engineering blogs or documentation from companies such as Google, Meta, Netflix, Stripe, Shopify, Airbnb, Uber, Vercel, Cloudflare, GitHub, and similar mature teams
• official framework or platform documentation
• postmortems or architecture writeups for comparable products

If exact FAANG-style analogues do not exist, use the closest high-quality engineering sources and say so explicitly.

Quality Standards

• Advocate: Specific evidence, not optimism. Acknowledge limitations while arguing strengths.
• Analyst: Genuinely neutral. Quantifiable comparisons. All major assumptions identified.
• Critic: Constructive, not destructive. Mitigations for every criticism. Prioritized by severity.
• All reviewers: Arguments must cite both codebase evidence and web-researched industry practice for similar features/applications.

Integration with Plan Mode

1. User enters plan mode → agent creates plan
2. /triumvirate → three subagents debate
3. Plan amended → user approves or iterates
4. Exit plan mode with final plan

Reference Files

• prompts/advocate.md — Full Advocate subagent prompt
• prompts/analyst.md — Full Analyst subagent prompt
• prompts/critic.md — Full Critic subagent prompt
• references/templates.md — Synthesis, amendment, and presentation templates
• references/example-output.md — Complete example of triumvirate output