Kefan-Lin/debate

Debate

Run a structured debate by dispatching subagents as debaters, moderating rounds between them, then synthesizing a comprehensive verdict. You are the moderator, not a debater — stay neutral until the synthesis step.

Quick start

User: "Debate whether we should rewrite the service in Rust."

1. Frame a sharp resolution + pick sides (default: 2 — Affirmative / Negative).
2. Round 1 — spawn each debater in parallel for opening statements.
3. Round 2+ — spawn each debater again, feeding them the opponent's last arguments.
4. Synthesize a verdict (you, or a dedicated Judge subagent).

Critical rule: subagents are stateless

Each Agent call starts cold with no memory of prior calls. A real debate therefore depends on you, the moderator, doing this every round:

• Keep the full running transcript yourself.
• When spawning a debater for a rebuttal, paste the opponent's latest arguments into its prompt. A debater that can't see what it's rebutting will monologue.
• Spawn same-round debaters in one message (parallel) so neither sees the other's current-round text — they react only to the previous round. This keeps it fair.
• Optimization — if your harness supports continuing a prior subagent (e.g. Claude Code returns an agentId / SendMessage), continue each debater instead of re-spawning: it already remembers its own prior argument, so you only need to paste the opponent's latest arguments. Still spawn same-round debaters separately to keep the round fair.

Workflow

Create a TodoWrite item per step for anything beyond a single quick round.

1. Frame — Restate the topic as a crisp, debatable resolution. Decide:

   • Sides: 2 (Affirmative/Negative) by default, or 3+ named perspectives for multi-stakeholder questions (e.g. "Build in-house" / "Buy SaaS" / "Adopt OSS"). See REFERENCE.md → Multi-perspective.
   • Rounds: 2 (opening + rebuttal) by default; 3 (add closing) for hard topics.
   • Fact-check: off by default; turn on for empirical / claim-heavy topics — a neutral verifier checks each side's factual claims before synthesis (step 4).
   • Judge: you synthesize by default; spawn a dedicated Judge subagent for maximal impartiality or a scored decision (step 5).
   • Confirm framing with the user only if the topic is ambiguous; otherwise proceed.

2. Round 1 — Opening statements — In a single message, spawn one subagent_type: "general-purpose" agent per side / perspective. Give each: the resolution, its assigned stance, the names of the other stances (so it knows the field), and the rules (evidence-based, steelman your side, concede nothing unearned). Collect every opening statement into your transcript.

3. Rebuttal round(s) — In a single message, re-spawn each debater. Each prompt now includes: the resolution, its stance, its own prior argument, and every other side's full prior arguments to rebut (for 3+ sides, paste all of them, labeled). Optionally add a closing round.

4. (Optional) Fact-check — If enabled, spawn one neutral verifier subagent with every side's factual claims pulled from the transcript. It marks each claim Supported / Disputed / Unverifiable with a one-line reason and does not take a side. Carry its findings into step 5 and discount flagged claims.

5. Synthesize the verdict — Reason over the transcript yourself, or spawn a dedicated Judge subagent with the complete transcript (plus any fact-check findings). Produce the structured verdict below.

See REFERENCE.md for copy-paste prompt templates for every role — opening, rebuttal, closing, fact-checker, judge — and modes.

Verdict format

Output a comprehensive synthesis — not just "who won":

• Resolution — the exact question debated.
• Strongest case per side — each position steelmanned in 2–4 bullets.
• Cruxes — the points of genuine disagreement that actually decide the question.
• Concessions & weak points — where each side gave ground or argued poorly.
• What would resolve it — the evidence, data, or test that would settle each crux.
• Verdict — a reasoned conclusion with an explicit confidence level (and the conditions under which it would flip). If genuinely balanced, say so and explain why.
• Caveats — assumptions, missing information, scope limits.

With 3+ perspectives, the Cruxes section maps which stances conflict and which are reconcilable, and the Verdict may recommend a hybrid or sequenced position rather than crowning one winner. If fact-check ran, list disputed/unverifiable claims and discount any conclusion that leaned on them.

Tuning

• Quick take: 2 sides, 1 round, you judge. Fast pro/con.
• Deep dive: 2–3 sides, 3 rounds, dedicated Judge subagent.
• Multi-perspective: 3+ named stances for multi-stakeholder calls; the verdict maps conflicts vs. reconcilable positions instead of picking a single winner.
• Red team: 1 proposal + 1 adversary whose only job is to break it.
• Fact-checked debate: enable the fact-checker step for empirical / claim-heavy topics so the verdict rests only on verified claims.