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

• Claude Code (CLI or VS Code extension)
• Codex CLI — npm install -g @openai/codex, then codex login
• Codex plugin for Claude Code — install from within Claude Code

Included skills

Skill files:

• skills/strategy-debate/SKILL.md
• skills/creator-critic/SKILL.md
• skills/options-challenge/SKILL.md

When To Use Which Skill

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:.

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

• Skills follow the language of the user brief unless the prompt explicitly asks for another output language.
• In strategy-debate, Protocol Summary follows the language of Topic. If the topic is mixed-language, the skill uses the main language of the brief.

Installation

Option 1: Claude Code plugin (recommended)

Install directly as a Claude Code plugin:

claude plugin install --source https://github.com/biyachuev/claude-debate-skills.git

Or add to a marketplace and install by name. See Claude Code plugin docs for details.

Option 2: global install script

Keep this repository anywhere on disk and run:

git clone https://github.com/biyachuev/claude-debate-skills.git
cd claude-debate-skills
./install.sh

This creates symlinks in ~/.claude/skills/ for all three skills.

To remove them later:

./uninstall.sh

Option 3: manual global symlinks

mkdir -p ~/.claude/skills
ln -s "/absolute/path/to/repo/skills/strategy-debate" ~/.claude/skills/strategy-debate
ln -s "/absolute/path/to/repo/skills/creator-critic" ~/.claude/skills/creator-critic
ln -s "/absolute/path/to/repo/skills/options-challenge" ~/.claude/skills/options-challenge

After that, the slash commands are available in any project opened in Claude Code.

Option 4: project-local skills

Copy or symlink the skill directories into the target project's .claude/skills/ folder.

mkdir -p /path/to/project/.claude/skills
ln -s "/absolute/path/to/repo/skills/strategy-debate" /path/to/project/.claude/skills/strategy-debate

This is useful if you want different projects to expose different skills.

Prompt Templates

If you prefer copy-paste prompts over slash commands, ready-made templates are available in two languages:

Troubleshooting

• Claude refuses to call Codex: make sure the Codex plugin is installed in Claude Code and codex login has completed successfully.
• The slash command does not exist: verify that the symlink exists under ~/.claude/skills/ and points to this repository.
• The skill works but no log file appears: check whether you passed No Log or Log File: none, and remember that relative log paths resolve from the currently open project.
• I only see Claude output: the skill is designed to stop instead of fabricating Codex output when Codex is unavailable.
• The symlink already exists: remove the old path manually or run ./uninstall.sh first, then install again.

Contributing

See CONTRIBUTING.md for guidelines on adding new skills or improving existing ones.

License

MIT

На русском

Набор переиспользуемых скиллов для Claude Code, которые превращают AI-советчиков в структурированных оппонентов.

Две модели, Claude (Anthropic) и Codex (OpenAI), спорят друг с другом по заданному протоколу вместо того, чтобы давать один бесспорный ответ. Разные данные обучения, разные слепые зоны, структурированный спор вскрывает то, что одна модель пропустила бы.

Три скилла

• /strategy-debate — глубокий анализ. Обе модели независимо формулируют позицию, в раундах перекрёстной критики атакуют слабые места друг друга, пересматривают аргументы. Финализатор подводит итог. Для бизнес-стратегий, архитектурных решений, планирования.
• /creator-critic — генерация и отсев. Одна модель предлагает 3-5 вариантов, другая разбирает каждый: ценность, главный изъян, скрытое допущение, вероятный сценарий провала. Creator пересматривает список и сужает до лучших. Поддерживает режим без Codex через Creator: Claude. Critic: Claude..
• /options-challenge — выбор из нескольких путей. Claude формулирует 2-4 варианта, Codex атакует каждый по trade-offs и рискам исполнения, Claude пересобирает рейтинг. Codex дожимает два лучших, и на выходе получается карта решения с явными компромиссами.

Быстрый старт

# 1. Установить Codex CLI
npm install -g @openai/codex && codex login

# 2. Установить Codex-плагин для Claude Code
/install-plugin codex

# 3. Установить debate skills как плагин
claude plugin install --source https://github.com/biyachuev/claude-debate-skills.git

# 4. Открыть любой проект в Claude Code и запустить
/strategy-debate Topic: Стоит ли выходить на SMB или enterprise первым? Constraints: команда из 6, нужна быстрая обратная связь.

Альтернатива: git clone + ./install.sh для установки через симлинки.

Что нужно

• Claude Code (CLI или расширение VS Code)
• Codex CLI — npm install -g @openai/codex, затем codex login
• Codex-плагин для Claude Code

Что важно знать

• Бриф может быть отправлен сразу двум вендорам: Anthropic и OpenAI.
• Если логирование включено, транскрипт дебатов сохраняется локально в Markdown.
• Чтобы не писать лог, используйте No Log, Log: none или Log File: none.
• Не прогоняйте через эти скиллы секреты, ключи, персональные данные и другой чувствительный контент.