Multi-perspective analysis powered by a team of independent agents. One is always Codex (different model family) for genuine diversity.
claude skill install hksw-io/debate-skill
/debate Should we migrate from REST to GraphQL?
/debate size:5 depth:deep Build vs buy for notifications
/debate Why is checkout dropping 30% at payment?| Arg | Example | Description |
|---|---|---|
size:N |
size:5 |
Panel size (3-7) |
depth:X |
depth:deep |
quick (1 round), standard (2-3), deep (4+) |
| (rest) | Should we migrate? |
Topic (required) |
- Context — explores codebase (or skips for non-code topics), asks high-impact questions if needed
- Complexity — scores topic on 3 dimensions (3-9), sets panel size and Codex reasoning effort
- Perspectives — derives domain-specific roles from topic (e.g., "SwiftData Expert", "QA Lead", "Product Owner"), not generic archetypes
- Team — spawns one agent per perspective, all working in parallel
- Rounds — thesis -> antithesis -> synthesis, with full history passed each round
- Report — decision, key trade-offs with consensus labels, recommendations, open questions (+ conditional: surprise finding, strongest dissent, cross-model divergence)
- Always a team — independent agents, not one model playing all roles
- Codex is a first-class member — same task list, messaging, and coordination as Claude agents
- Domain roles — "Core Data Advocate" not "Contrarian"
- Platform quality — auto-detects Apple/Microsoft/Android/Web and adds a platform-native quality perspective
- Cumulative context — every round gets full debate history, critical for stateless Codex agent
- Evidence tagging — claims tagged [grounded], [informed], or [speculative]
- Mandatory question enforcement — agent questions surfaced to user via AskUserQuestion, not silently skipped
- Codex CLI installed
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1in Claude Code settings
MIT