workflow-debate

Research backing: MIT studies show multi-perspective AI deliberation improves decision accuracy by 7-45%. The AI Council Framework and Debate-Based Consensus patterns demonstrate that structured disagreement surfaces blind spots that single-perspective analysis misses.

Economy mode (-e): no prerequisites, uses sequential subagents instead of Agent Teams.

<quick_start>

Minimal council (3 steps):

1. /council Should we migrate our monolith to microservices?
2. Councillors debate across 2 rounds (default)
3. Receive structured verdict: PROCEED / PROCEED WITH CONDITIONS / RECONSIDER / DO NOT PROCEED

With options:

/council -c 5 -r 3 -s -f security Our plan to store PII in a third-party service

</quick_start>

Flags:

Councillor assignment by -c N:

Parse at start:

1. Check for -r N → set {rounds} (default 2, clamp 1-4)
2. Check for -c N → set {councillor_count} (default 4, clamp 3-6)
3. Check for -s → set {save_mode} = true
4. Check for -e → set {economy_mode} = true
5. Check for -f X → set {focus} = X
6. Validate: if -c 6 without -f → warn and reduce to 5
7. Validate: if -e and Agent Teams env var set → note economy override
8. Remove all flags from input → store as {topic}

<state_variables> Capture at start and persist throughout all steps:

• {topic} - The idea/plan to evaluate (flags removed)
• {rounds} - Number of debate rounds (1-4, default 2)
• {councillor_count} - Number of councillors (3-6, default 4)
• {save_mode} - true if -s flag passed
• {economy_mode} - true if -e flag passed
• {focus} - Domain focus if -f passed (technical/business/UX/security/ethics/custom)
• {councillors} - List of active councillor roles
• {team_name} - Team name (if Agent Teams mode)
• {output_dir} - Path to output directory (if save_mode)
• {opening_statements} - Collected opening positions from step-02
• {debate_log} - Accumulated debate exchanges from step-03
• {position_evolution} - How each councillor's position changed across rounds </state_variables>

Summary:

Phase 1: Setup (step-01)

1. Parse flags and validate configuration
2. Determine councillor roster based on -c N
3. Create team (Agent Teams) or prepare sequential execution (economy)
4. Spawn all councillors in parallel

Phase 2: Opening Statements (step-02)

5. Each councillor independently analyzes the topic from their perspective
6. No cross-contamination — councillors don't see each other's positions yet
7. Moderator collects all opening statements

Phase 3: Debate Rounds (step-03)

8. Share all positions with every councillor
9. Each councillor must challenge 2+ specific points from others
10. Peer exchange: councillors message each other directly via SendMessage
11. Moderator injects anti-groupthink questions between rounds
12. Track position evolution across rounds
13. Repeat for {rounds} rounds

Phase 4: Verdict (step-04)

14. Synthesize consensus and dissensus points
15. Build risk assessment matrix
16. Produce structured verdict with confidence level
17. Present strongest counter-argument (steelmanned)
18. Save transcript (if -s) and cleanup team

<step_files> Each step is a separate file for progressive context loading:

• steps/step-01-setup.md - Parse flags, create team, spawn councillors
• steps/step-02-opening.md - Independent opening statements (no cross-contamination)
• steps/step-03-debate.md - Cross-examination rounds + peer exchange
• steps/step-04-verdict.md - Synthesis, verdict, cleanup </step_files>

<entry_point> <step_0 name="Initialize">

FIRST ACTION - Parse flags:

1. Check for -r N → set {rounds} (default 2, validate 1-4)
2. Check for -c N → set {councillor_count} (default 4, validate 3-6)
3. Check for -s → set {save_mode} = true
4. Check for -e → set {economy_mode} = true
5. Check for -f X → set {focus} = X
6. Validate flag combinations (see parameters section)
7. Remove flags from input → store as {topic}

THEN: Load steps/step-01-setup.md to begin council setup. </step_0> </entry_point>

<execution_rules>

• Load one step at a time - Only load the current step file
• Persist state variables across all steps
• Follow next_step directive at end of each step
• You are the Moderator - Orchestrate, don't argue. Stay neutral.
• Economy mode - Use sequential Task calls with subagent_type Plan instead of Agent Teams
• Peer exchange is the differentiator - In standard mode, councillors MUST message each other directly
• Anti-groupthink - If positions converge too quickly, inject challenging questions
• Steelman, don't strawman - Present the strongest version of opposing arguments
• Read-only councillors - Councillors analyze and argue, they NEVER modify files
• Always cleanup - Shutdown all councillors + TeamDelete when done

</execution_rules>

<success_criteria>

• All councillors produced opening statements independently
• At least {rounds} rounds of cross-examination completed
• Peer exchange occurred (councillors challenged each other directly)
• Position evolution tracked across rounds
• Structured verdict produced with confidence level
• Consensus AND dissensus points identified
• Risk assessment matrix included
• Strongest counter-argument steelmanned
• Team cleaned up properly (shutdown + delete)
• Transcript saved (if -s flag)

</success_criteria>