Deliberative Refinement

Stop asking AI for answers. Make it defend them.

A structured multi-agent deliberation framework for complex reasoning, uncertain situations, and consequential decisions.

What Is This?

Deliberative Refinement forces AI through multiple rounds of structured critique—different expert perspectives attack from different angles, web search grounds claims between passes, iteration continues until only ideas that survive adversarial pressure remain.

The Core Idea

Instead of prompt → output → polish, you architect critique rounds:

1. Expert councils attack the draft from different angles (code reviewers, strategy councils, elimination tournaments)
2. Web search grounds claims with actual evidence between rounds
3. Iteration continues until weak ideas die and strong ones survive

The Breakthrough

Six months ago, this required orchestrating a dozen different models—one for generation, another for critique, a third for fact-checking. Tool-interspersed reasoning collapsed all of that into a single model that switches roles on demand. Same quality, fraction of the complexity.

flowchart LR subgraph Input A[📄 Draft] end

subgraph "Phase 0: Intent Detection"
    B{Detect Mode}
    B -->|CREATE| C1[Generate First]
    B -->|REFINE| C2[Analyze Input]
    B -->|DECIDE| C3[Frame Options]
    B -->|DISCOVER| C4[Explore Space]
end

subgraph "Phase 1: Council Selection"
    D{Select Council}
    D -->|Code/Tech| E1[Structured Review]
    D -->|A vs B| E2[Elimination Tournament]
    D -->|Math/Logic| E3[Meta-Reasoning]
    D -->|General| E4[Expert Council]
end

subgraph "Phase 2-3: Deliberation Rounds"
    F[🔍 Probe 1\nWeb Grounding]
    G[⚔️ Round 1\nX Agents Deliberate]
    H[🔍 Probe 2\nEvidence Check]
    I[⚔️ Round 2\nPositions Refined]
    J[🔍 Probe 3\nFinal Grounding]
    K[⚔️ Round 3\nConvergence Check]
end

subgraph "Phase 4: Output"
    L{Converged?}
    L -->|Yes| M[✅ Final Output]
    L -->|No + Echo Chamber| N[Add Adversarial Agent]
    N --> G
end

A --> B
C1 --> D
C2 --> D
C3 --> D
C4 --> D
E1 --> F
E2 --> F
E3 --> F
E4 --> F
F --> G
G --> H
H --> I
I --> J
J --> K
K --> L

Quick Start

As a Claude Skill

# Clone this repo
git clone https://github.com/YOUR_USERNAME/deliberative-refinement.git

# Copy to your Claude skills directory
cp -r deliberative-refinement ~/.claude/skills/

Or install via marketplace:

/plugin marketplace add YOUR_USERNAME/deliberative-refinement

As a Slash Command

Copy slashcommand-deliberative-refinement.md to your preferred slash command directory.

How It Works

The Formula

V(X, Y, S) = X agents, Y rounds, S probes between rounds
Default: V(8, 3, 1) = 8 agents × 3 rounds × 1 web probe per gap

Profiles

Profile	V(X,Y,S)	Use Case
Lite	V(3,1,0)	Quick sanity check
Standard	V(8,3,1)	Default
Deep	V(12,5,2)	High-stakes decisions
Exhaustive	V(15,5,3)	Critical outcomes

Modes

Mode	Trigger
CREATE	No input → Generate → Validate
REFINE	Input provided → Deliberate → Return
DECIDE	Options presented → Evaluate → Recommend
DISCOVER	Approach unclear → Explore → Synthesize

Council Types

Council	Best For
Expert Council	General problems, synthesis
Structured Review	Code, technical analysis
Elimination Tournament	Binary A vs B decisions
Meta-Reasoning	Math, proofs, sequential logic
Parallel Groups	Large scope, budget limits

Why This Matters

Single-pass prompting optimizes for plausible.

Deliberative refinement optimizes for robust.

It's the difference between "AI said so" and "survived peer review."

Use Cases

• Strategy docs that survive executive review
• Technical specs that don't detonate in production
• Research that holds up to scrutiny
• Legal documents that don't need 3 revision cycles
• Architecture decisions that don't implode at scale

File Structure

deliberative-refinement/
├── SKILL.md                              # Claude skill definition
├── slashcommand-deliberative-refinement.md  # Slash command version
├── marketplace.json                      # For marketplace installation
├── references/
│   ├── councils.md                       # Council type specifications
│   └── profiles.md                       # Intensity profiles
└── README.md                             # This file

Contributing

PRs welcome. If you've found improvements to council configurations, probe strategies, or convergence criteria, open an issue or submit a PR.

License

MIT

---

Generation is table stakes. Making AI defend its ideas until they break or bend? That's the new standard.