Claude Debate Skills
Reusable Claude Code skills for strategy discussions, idea generation, critique, and decision pressure-testing.
Two AI models, Claude (Anthropic) and Codex (OpenAI), debate each other through structured protocols instead of giving you a single unchallenged answer. Different training data, different blind spots, structured disagreement that surfaces what one model alone would miss.
Quick Start
# 1. Install Codex CLI
npm install -g @openai/codex && codex login
# 2. Install Codex plugin for Claude Code (from within Claude Code)
/install-plugin codex
# 3. Install debate skills as a plugin
claude plugin install --source https://github.com/biyachuev/claude-debate-skills.git
# 4. Open any project in Claude Code and run:
/strategy-debate Topic: Should we go after SMB or enterprise first? Constraints: team of 6, need fast feedback loops.
Alternative: clone the repository and use ./install.sh for symlink-based installation (see Installation below).
Requirements
Included skills
| Skill | Purpose | Solo fallback |
|---|
strategy-debate | Two-voice debate with cross-critique, revision, and final synthesis | No — two-voice by design |
creator-critic | One side generates options, the other pressure-tests them, then the creator revises | Yes — pass Creator: Claude. Critic: Claude. |
options-challenge | Compare several strategic paths, challenge trade-offs, and rank the best option | No — Codex challenge is structural |
Skill files:
When To Use Which Skill
| Situation | Recommended skill | Why |
|---|
| You need a hard comparison between several plausible paths | options-challenge | Forces ranking and trade-offs instead of a soft brainstorm |
| You need two independent positions and cross-critique before a decision | strategy-debate | Best for strategy, architecture, research direction, and planning |
| You need breadth first, then filtering and revision | creator-critic | Good for ideas, naming, positioning, campaigns, and concept generation |
When Debate Is A Bad Fit
- The answer is mostly factual and can be verified directly.
- The task is trivial and not worth spending tokens from two models.
- You need the fastest possible answer.
- The brief contains sensitive material that you should not send to multiple vendors.
Invocation Examples
/strategy-debate Finalizer: Claude. Log File: smb-vs-enterprise. Topic: Should we go after SMB or enterprise first? Constraints: team of 6, low sales capacity, need fast feedback loops.
/creator-critic Log File: pricing-ideas. Brief: Generate and pressure-test pricing ideas for an AI debate copilot. Constraints: simple plans, low trust, early-stage product.
/creator-critic Creator: Claude. Critic: Claude. Finalizer: Claude. Log File: pricing-ideas-solo. Brief: Generate and pressure-test pricing ideas for an AI debate copilot. Constraints: simple plans, low trust, early-stage product.
/options-challenge Log File: gtm-paths-q3. Decision Brief: Compare direct-to-student, educator-led, and partnership GTM paths. Constraints: small team, one quarter to show traction.
Log Files
All skills support Log File: or Save As:.
| Input | Result |
|---|
Log File: my-note | .claude/debate-logs/my-note.md |
Log File: my-note.md | .claude/debate-logs/my-note.md |
Log File: notes/decisions/my-note.md | Relative path in the current project |
| Omitted | Session-based filename in .claude/debate-logs/ |
No Log / Log: none / Log File: none | No file written, full debate in chat only |
When skills are installed globally, the log file is resolved relative to the currently active project, not this repository.
Privacy And Data Handling
- Your brief may be sent to both Claude and Codex during the debate.
- If logging is enabled, the visible transcript is written to a local Markdown file.
- Use
No Log, Log: none, or Log File: none if you do not want a local transcript.
- Do not use these skills for credentials, secrets, regulated data, or anything you would not want to send to multiple AI providers.
Language Behavior
