adversarial-spec

A Claude Code plugin that iteratively refines product specifications through multi-model debate until consensus is reached.

Key insight: A single LLM reviewing a spec will miss things. Multiple LLMs debating a spec will catch gaps, challenge assumptions, and surface edge cases that any one model would overlook. The result is a document that has survived rigorous adversarial review.

Claude is an active participant, not just an orchestrator. Claude provides independent critiques, challenges opponent models, and contributes substantive improvements alongside external models.

Quick Start

# 1. Add the marketplace and install the plugin
claude plugin marketplace add zscole/adversarial-spec
claude plugin install adversarial-spec

# 2. Set at least one API key
export OPENAI_API_KEY="sk-..."

# 3. Run it
/adversarial-spec "Build a rate limiter service with Redis backend"

How It Works

You describe product --> Claude drafts spec --> Multiple LLMs critique in parallel
        |                                              |
        |                                              v
        |                              Claude synthesizes + adds own critique
        |                                              |
        |                                              v
        |                              Revise and repeat until ALL agree
        |                                              |
        +--------------------------------------------->|
                                                       v
                                            User review period
                                                       |
                                                       v
                                            Final document output

1. Describe your product concept or provide an existing document
2. (Optional) Start with an in-depth interview to capture requirements
3. Claude drafts the initial document (PRD or tech spec)
4. Document is sent to opponent models (GPT, Gemini, Grok, etc.) for parallel critique
5. Claude provides independent critique alongside opponent feedback
6. Claude synthesizes all feedback and revises
7. Loop continues until ALL models AND Claude agree
8. User review period: request changes or run additional cycles
9. Final converged document is output

Requirements

• Python 3.10+
• litellm package: pip install litellm
• API key for at least one LLM provider

Supported Models

Provider	Env Var	Example Models
OpenAI	OPENAI_API_KEY	gpt-4o, gpt-4-turbo, o1
Google	GEMINI_API_KEY	gemini/gemini-2.0-flash, gemini/gemini-pro
xAI	XAI_API_KEY	xai/grok-3, xai/grok-beta
Mistral	MISTRAL_API_KEY	mistral/mistral-large, mistral/codestral
Groq	GROQ_API_KEY	groq/llama-3.3-70b-versatile
Deepseek	DEEPSEEK_API_KEY	deepseek/deepseek-chat

Check which keys are configured:

python3 ~/.claude/skills/adversarial-spec/scripts/debate.py providers

AWS Bedrock Support

For enterprise users who need to route all model calls through AWS Bedrock (e.g., for security compliance or inference gateway requirements):

# Enable Bedrock mode
python3 ~/.claude/skills/adversarial-spec/scripts/debate.py bedrock enable --region us-east-1

# Add models enabled in your Bedrock account
python3 ~/.claude/skills/adversarial-spec/scripts/debate.py bedrock add-model claude-3-sonnet
python3 ~/.claude/skills/adversarial-spec/scripts/debate.py bedrock add-model claude-3-haiku

# Check configuration
python3 ~/.claude/skills/adversarial-spec/scripts/debate.py bedrock status

# Disable Bedrock mode
python3 ~/.claude/skills/adversarial-spec/scripts/debate.py bedrock disable

When Bedrock is enabled, all model calls route through Bedrock - no direct API calls are made. Use friendly names like claude-3-sonnet which are automatically mapped to Bedrock model IDs.

Configuration is stored at ~/.claude/adversarial-spec/config.json.

Usage

Start from scratch:

/adversarial-spec "Build a rate limiter service with Redis backend"

Refine an existing document:

/adversarial-spec ./docs/my-spec.md

You will be prompted for:

1. Document type: PRD (business/product focus) or tech spec (engineering focus)
2. Interview mode: Optional in-depth requirements gathering session
3. Opponent models: Comma-separated list (e.g., gpt-4o,gemini/gemini-2.0-flash,xai/grok-3)

More models = more perspectives = stricter convergence.

Document Types

PRD (Product Requirements Document)

For stakeholders, PMs, and designers.

Sections: Executive Summary, Problem Statement, Target Users/Personas, User Stories, Functional Requirements, Non-Functional Requirements, Success Metrics, Scope (In/Out), Dependencies, Risks

Critique focuses on: Clear problem definition, well-defined personas, measurable success criteria, explicit scope boundaries, no technical implementation details

Technical Specification

For developers and architects.

Sections: Overview, Goals/Non-Goals, System Architecture, Component Design, API Design (full schemas), Data Models, Infrastructure, Security, Error Handling, Performance/SLAs, Observability, Testing Strategy, Deployment Strategy

Critique focuses on: Complete API contracts, data model coverage, security threat mitigation, error handling, specific performance targets, no ambiguity for engineers

Core Features

Interview Mode

Before the debate begins, opt into an in-depth interview session to capture requirements upfront.

Covers: Problem context, users/stakeholders, functional requirements, technical constraints, UI/UX, tradeoffs, risks, success criteria

The interview uses probing follow-up questions and challenges assumptions. After completion, Claude synthesizes answers into a complete spec before starting the adversarial debate.

Claude's Active Participation

Each round, Claude:

1. Reviews opponent critiques for validity
2. Provides independent critique (what did opponents miss?)
3. States agreement/disagreement with specific points
4. Synthesizes all feedback into revisions

Display format:

--- Round N ---
Opponent Models:
- [GPT-4o]: critiqued: missing rate limit config
- [Gemini]: agreed

Claude's Critique:
Security section lacks input validation strategy. Adding OWASP top 10 coverage.

Synthesis:
- Accepted from GPT-4o: rate limit configuration
- Added by Claude: input validation, OWASP coverage
- Rejected: none

Early Agreement Verification

If a model agrees within the first 2 rounds, Claude is skeptical. The model is pressed to:

• Confirm it read the entire document
• List specific sections reviewed
• Explain why it agrees
• Identify any remaining concerns

This prevents false convergence from models that rubber-stamp without thorough review.

User Review Period

After all models agree, you enter a review period with three options:

1. Accept as-is: Document is complete
2. Request changes: Claude updates the spec, you iterate without a full debate cycle
3. Run another cycle: Send the updated spec through another adversarial debate

Additional Review Cycles

Run multiple cycles with different strategies:

• First cycle with fast models (gpt-4o), second with stronger models (o1)
• First cycle for structure/completeness, second for security focus
• Fresh perspective after user-requested changes

PRD to Tech Spec Flow

When a PRD reaches consensus, you're offered the option to continue directly into a Technical Specification based on the PRD. This creates a complete documentation pair in a single session.

Advanced Features

Critique Focus Modes

Direct models to prioritize specific concerns:

--focus security      # Auth, input validation, encryption, vulnerabilities
--focus scalability   # Horizontal scaling, sharding, caching, capacity
--focus performance   # Latency targets, throughput, query optimization
--focus ux            # User journeys, error states, accessibility
--focus reliability   # Failure modes, circuit breakers, disaster recovery
--focus cost          # Infrastructure costs, resource efficiency

Model Personas

Have models critique from specific professional perspectives:

--persona security-engineer      # Thinks like an attacker
--persona oncall-engineer        # Cares about debugging at 3am
--persona junior-developer       # Flags ambiguity and tribal knowledge
--persona qa-engineer            # Missing test scenarios
--persona site-reliability       # Deployment, monitoring, incidents
--persona product-manager        # User value, success metrics
--persona data-engineer          # Data models, ETL implications
--persona mobile-developer       # API design for mobile
--persona accessibility-specialist  # WCAG, screen readers
--persona legal-compliance       # GDPR, CCPA, regulatory

Custom personas also work: --persona "fintech compliance officer"

Context Injection

Include existing documents for models to consider:

--context ./existing-api.md --context ./schema.sql

Use cases:

• Existing API documentation the new spec must integrate with
• Database schemas the spec must work with
• Design documents or prior specs for consistency
• Compliance requirements documents

Session Persistence and Resume

Long debates can crash or need to pause. Sessions save state automatically:

# Start a named session
echo "spec" | python3 debate.py critique --models gpt-4o --session my-feature-spec

# Resume where you left off
python3 debate.py critique --resume my-feature-spec

# List all sessions
python3 debate.py sessions

Sessions save:

• Current spec state
• Round number
• All configuration (models, focus, persona, etc.)
• History of previous rounds

Sessions are stored in ~/.config/adversarial-spec/sessions/.

Auto-Checkpointing

When using sessions, each round's spec is saved to .adversarial-spec-checkpoints/:

.adversarial-spec-checkpoints/
├── my-feature-spec-round-1.md
├── my-feature-spec-round-2.md
└── my-feature-spec-round-3.md

Use these to rollback if a revision makes things worse.

Preserve Intent Mode

Convergence can sand off novel ideas when models interpret "unusual" as "wrong". The --preserve-intent flag makes removal expensive:

--preserve-intent

When enabled, models must:

1. Quote exactly what they want to remove or substantially change
2. Justify the harm - not just "unnecessary" but what concrete problem it causes
3. Distinguish error from preference - only remove things that are factually wrong, contradictory, or risky
4. Ask before removing unusual but functional choices: "Was this intentional?"

This shifts the default from "sand off anything unusual" to "add protective detail while preserving distinctive choices."

Use when:

• Your spec contains intentional unconventional choices
• You want models to challenge your ideas, not homogenize them
• Previous rounds removed things you wanted to keep

Cost Tracking

Every critique round displays token usage and estimated cost:

=== Cost Summary ===
Total tokens: 12,543 in / 3,221 out
Total cost: $0.0847

By model:
  gpt-4o: $0.0523 (8,234 in / 2,100 out)
  gemini/gemini-2.0-flash: $0.0324 (4,309 in / 1,121 out)

Saved Profiles

Save frequently used configurations:

# Create a profile
python3 debate.py save-profile strict-security \
  --models gpt-4o,gemini/gemini-2.0-flash \
  --focus security \
  --doc-type tech

# Use a profile
python3 debate.py critique --profile strict-security < spec.md

# List profiles
python3 debate.py profiles

Profiles are stored in ~/.config/adversarial-spec/profiles/.

Diff Between Rounds

See exactly what changed between spec versions:

python3 debate.py diff --previous round1.md --current round2.md

Export to Task List

Extract actionable tasks from a finalized spec:

cat spec-output.md | python3 debate.py export-tasks --models gpt-4o --doc-type prd

Output includes title, type, priority, description, and acceptance criteria.

Use --json for structured output suitable for importing into issue trackers.

Telegram Integration (Optional)

Get notified on your phone and inject feedback during the debate.

Setup:

1. Message @BotFather on Telegram, send /newbot, follow prompts
2. Copy the bot token
3. Run: python3 ~/.claude/skills/adversarial-spec/scripts/telegram_bot.py setup
4. Message your bot, run setup again to get your chat ID
5. Set environment variables:

export TELEGRAM_BOT_TOKEN="..."
export TELEGRAM_CHAT_ID="..."

Features:

• Async notifications when rounds complete (includes cost)
• 60-second window to reply with feedback (incorporated into next round)
• Final document sent to Telegram when debate concludes

Output

Final document is:

• Complete, following full structure for document type
• Vetted by all models until unanimous agreement
• Ready for stakeholders without further editing

Output locations:

• Printed to terminal
• Written to spec-output.md (PRD) or tech-spec-output.md (tech spec)
• Sent to Telegram (if enabled)

Debate summary includes rounds completed, cycles run, models involved, Claude's contributions, cost, and key refinements made.

CLI Reference

# Core commands
debate.py critique --models MODEL_LIST --doc-type TYPE [OPTIONS] < spec.md
debate.py critique --resume SESSION_ID
debate.py diff --previous OLD.md --current NEW.md
debate.py export-tasks --models MODEL --doc-type TYPE [--json] < spec.md

# Info commands
debate.py providers      # List providers and API key status
debate.py focus-areas    # List focus areas
debate.py personas       # List personas
debate.py profiles       # List saved profiles
debate.py sessions       # List saved sessions

# Profile management
debate.py save-profile NAME --models ... [--focus ...] [--persona ...]

# Bedrock management
debate.py bedrock status                      # Show Bedrock configuration
debate.py bedrock enable --region REGION      # Enable Bedrock mode
debate.py bedrock disable                     # Disable Bedrock mode
debate.py bedrock add-model MODEL             # Add model to available list
debate.py bedrock remove-model MODEL          # Remove model from list
debate.py bedrock list-models                 # List built-in model mappings

Options:

• --models, -m - Comma-separated model list
• --doc-type, -d - prd or tech
• --focus, -f - Focus area (security, scalability, performance, ux, reliability, cost)
• --persona - Professional persona
• --context, -c - Context file (repeatable)
• --profile - Load saved profile
• --preserve-intent - Require justification for removals
• --session, -s - Session ID for persistence and checkpointing
• --resume - Resume a previous session
• --press, -p - Anti-laziness check
• --telegram, -t - Enable Telegram
• --json, -j - JSON output

File Structure

adversarial-spec/
├── .claude-plugin/
│   └── plugin.json           # Plugin metadata
├── README.md
├── LICENSE
└── skills/
    └── adversarial-spec/
        ├── SKILL.md          # Skill definition and process
        └── scripts/
            ├── debate.py     # Multi-model debate orchestration
            └── telegram_bot.py   # Telegram notifications

License

MIT