A markdown-driven decision governance system using a virtual board of AI agents.
The goal is not agreement. The goal is durable, falsifiable, and reviewable decisions.
Conclave is a structured decision-making framework that uses AI agents acting as specialized board members to stress-test important decisions before you make them.
Instead of making decisions in isolation or seeking validation from agreeable sources, Conclave forces you to:
- Articulate your reasoning clearly
- Face adversarial perspectives from multiple angles
- Consider failure modes before commitment
- Create falsifiable predictions
- Build a reviewable decision history
This framework is designed for founders, product leaders, and decision makers who need cognitive friction, not cheerleading.
Use Conclave when a decision:
- Introduces a new core mechanism
- Redistributes power or money
- Reduces reversibility
- Locks in platform identity or narrative
- Commits sustained resources
- Produces emotional relief or urgency
If uncertain, assume it applies.
This is not for trivial decisions. This is for decisions where being wrong is expensive and where momentum can override judgment.
See WORKFLOW.md for a visual diagram.
- Write a Board Memo — Articulate the decision, context, options, and your preferred direction
- Collect Role Feedback — Each AI board member (5 roles) evaluates from their mandate
- Synthesize Feedback — Compile all perspectives without resolving disagreements
- Run Pre-Mortem (if required) — Assume the decision failed 18-36 months later and work backward
- Apply Temporal Distance — Wait 24-72 hours without reviewing materials
- Re-read Cold — Approach all feedback with fresh perspective
- Make Decision — Now, and only now, decide
- Document — Write the authoritative decision record
Each role has a distinct mandate and is explicitly skeptical of different failure modes:
- Chair / Governance — Preserves mission coherence, challenges drift
- Economics & Incentives — Stress-tests for gaming and exploitation
- Technical & Scalability — Ensures feasibility and maintainability
- Product & User Advocate — Represents lived user experience
- External Threats — Maps competitive and strategic risks
No role sees another's feedback. Each acts independently.
- Access to an agentic AI system (Claude Code, Cursor, Windsurf, etc.)
- A decision worth the process
- Willingness to be challenged
-
Copy this repository
git clone https://github.com/mlazarev/conclave.git cd conclave -
⚠️ Delete the examples folderrm -rf examples/ # Prevents example context from polluting your decisions -
Create your context file
cp BOOTSTRAP_CONTEXT_TEMPLATE.md BOOTSTRAP_CONTEXT.md
Fill in
BOOTSTRAP_CONTEXT.mdwith your organization's information -
Create a decision folder
mkdir -p decisions/2026-01-10_your-decision-slug
-
Run the decision process Point your AI agent to the execution file:
Please follow RUN_DECISION.md to help me evaluate my decision about [topic] in decisions/2026-01-10_your-decision-slug/The agent will guide you through all 8 steps.
For step-by-step execution instructions, see RUN_DECISION.md
That file contains complete agent-executable instructions for all 8 steps. Simply point your AI agent to it:
Please follow RUN_DECISION.md to help me evaluate my decision
about [topic] in decisions/YYYY-MM-DD_slug/
- Write Board Memo — Agent helps you articulate the decision
- Collect Role Feedback — Agent acts as 5 board members independently
- Synthesize Feedback — Agent compiles all perspectives
- Determine Pre-Mortem Need — Agent evaluates if pre-mortem is required
- Run Pre-Mortem — Agent conducts failure analysis (if needed)
- Temporal Distance — You wait 24-72 hours (mandatory)
- Make Decision — You decide after cold re-read
- Write Decision Record — Agent helps document the decision
See WORKFLOW.md for a visual diagram of the process.
conclave/
├── README.md # This file
├── BOARD_CHARTER.md # Board purpose and operating principles
├── BOOTSTRAP_CONTEXT_TEMPLATE.md # Template for your organization context
├── BOOTSTRAP_CONTEXT.md # Your context (you create this)
├── AGENT_PROMPT.md # Instructions for AI board members
│
├── protocols/
│ ├── PROTOCOL_PRE_MORTEM.md # Pre-mortem process
│ └── PROTOCOL_TEMPORAL_DISTANCE.md # Mandatory delay protocol
│
├── roles/
│ ├── ROLE_CHAIR_GOVERNANCE.md # Mission & governance role
│ ├── ROLE_ECONOMICS_INCENTIVES.md # Economic skeptic role
│ ├── ROLE_TECHNICAL_REALIST.md # Technical feasibility role
│ ├── ROLE_PRODUCT_USER_ADVOCATE.md # User experience role
│ └── ROLE_EXTERNAL_THREATS.md # Competition & threats role
│
├── templates/
│ ├── TEMPLATE_BOARD_MEMO.md # For writing proposals
│ ├── TEMPLATE_BOARD_ROLE_FEEDBACK.md # For role responses
│ ├── TEMPLATE_BOARD_FEEDBACK_SUMMARY.md # For synthesis
│ ├── TEMPLATE_PRE_MORTEM_OUTPUT.md # For pre-mortem outputs
│ └── TEMPLATE_DECISION_RECORD.md # For final decisions
│
├── decisions/ # Your decision history
│ └── YYYY-MM-DD_decision-slug/
│ ├── 01_BOARD_MEMO.md
│ ├── 02_ROLE_FEEDBACK/
│ │ ├── chair.md
│ │ ├── economics.md
│ │ ├── technical.md
│ │ ├── product_user.md
│ │ └── external_threats.md
│ ├── 03_BOARD_FEEDBACK_SUMMARY.md
│ ├── 04_PRE_MORTEM/ # (if required)
│ │ ├── 00_SYNTHESIS.md
│ │ ├── chair.md
│ │ ├── economics.md
│ │ ├── technical.md
│ │ ├── product_user.md
│ │ └── external_threats.md
│ └── 05_DECISION_RECORD.md
│
└── examples/ # Sample materials
└── sample-decision/ # Complete example workflow
Conclave is designed to work seamlessly with agentic AI systems like:
- Claude Code (CLI)
- Cursor IDE
- Windsurf
- GitHub Copilot Workspace
- Any AI agent with file system access
Simply point your AI agent to RUN_DECISION.md:
Please follow RUN_DECISION.md to help me evaluate my decision
about [topic] in decisions/YYYY-MM-DD_slug/
The agent will:
- Read all instructions from RUN_DECISION.md
- Read all necessary context files automatically
- Execute all automatable steps (1-5, 8)
- Prompt you when human steps are required (6-7)
- Run board roles in parallel if supported (faster, more independent)
- Claude (Anthropic): Sonnet or Opus — best for adversarial reasoning
- GPT-4 (OpenAI): Good for structured feedback
- Other models: Any model with strong reasoning and instruction-following
- Point to RUN_DECISION.md — Don't paraphrase the instructions
- Let the agent drive — It knows what to do
- Don't cherry-pick outputs — Save all responses unchanged, even if harsh
- Trust the process — If all roles agree with you, something is wrong
- Complete temporal distance — Don't skip the mandatory 24-72 hour wait
You can customize Conclave for your specific needs:
- Modify roles — Edit role files to match your domain
- Add roles — Create new role files following the existing pattern
- Adjust verbosity — Change constraints in role definitions
- Modify templates — Adapt templates to your decision types
- Change thresholds — Adjust when pre-mortems or temporal distance apply
Copy an existing role file and modify:
- Role identity and mandate
- What it optimizes for
- What it's skeptical of
- What it explicitly ignores
- Required behaviors
- Verbosity constraints
Keep roles adversarial, not cooperative.
- Adapt: You're using this for the intended purpose (strategic decisions)
- Fork: You want to apply this to a completely different domain (e.g., personal decisions, academic review)
- Use this for decisions where being wrong is expensive
- Let roles be harsh — that's the point
- Wait the full temporal distance period
- Review decisions quarterly against outcomes
- Track which roles you tend to override
- Don't use this for trivial decisions (it's too heavy)
- Don't shop for validation — if you ignore all roles, ask why
- Don't skip the temporal distance step
- Don't revise board feedback to soften it
- Don't treat this as a rubber stamp
- You feel slightly uncomfortable with some feedback
- You discover assumptions you didn't know you had
- The pre-mortem surfaces a scenario that changes your confidence
- You catch yourself trying to justify rather than evaluate
- All roles agree with you
- Feedback feels generic or obvious
- You're frustrated the board "doesn't understand"
- You're rushing through steps to get to yes
Yes. Use it for decisions that matter.
That's fine — you decide, not them. But ask yourself: why did you run the process? Document your reasoning in the decision record.
You can, but it's usually a sign your board memo wasn't clear enough. Consider revising the memo and rerunning from scratch.
For major decisions, yes. For smaller decisions, you can run a subset, but document which roles you skipped and why.
With an agentic AI system:
- Steps 1-5, 8 (agent-executable): Minutes to complete, depending on your AI system
- Steps 6-7 (human-required): 24-72 hours for temporal distance + decision time
The mandatory delay is Step 6 (Temporal Distance): You must wait 24-72 hours between feedback and decision. This cannot be shortened or automated.
Total calendar time: 1-3 days (mostly the mandatory wait period)
AI agents can execute Steps 1-5 and 8 much faster than humans, and can run board roles in parallel for even greater speed.
Yes, but designate one person as the "decision maker" who owns the final call. Conclave is not consensus-building.
This framework does not replace professional advice. Use it to sharpen your thinking, then consult appropriate experts.
Your decision records can be public or private. The framework itself is open source.
Quarterly (or after 5-10 decisions), review:
- Falsification criteria — Were predictions accurate?
- Override patterns — Which roles do you consistently dismiss?
- Outcome alignment — Did pre-mortems predict actual failures?
- Process violations — Did you skip temporal distance or cherry-pick feedback?
The goal is pattern recognition, not perfect predictions.
See CONTRIBUTING.md for guidelines on:
- Suggesting role improvements
- Proposing new templates
- Sharing decision examples (anonymized)
- Reporting issues
This project is released under the MIT License. See LICENSE for details.
You are free to use, modify, and distribute this framework for any purpose.
This framework emerged from real decision-making struggles and draws on:
- Red team / blue team exercises
- Pre-mortem techniques (Gary Klein)
- Adversarial collaboration
- Falsification-first thinking (Popper)
The structure is opinionated by design. Friction is a feature.
- Issues: GitHub Issues
- Discussions: GitHub Discussions
- Examples: See
examples/directory
Remember: The goal is not agreement. The goal is durable, falsifiable, and reviewable decisions.
If this process feels comfortable, you're probably not using it correctly.