Council

Debate $ARGUMENTS from multiple perspectives:

Execution Contract (read first)

This skill is a delegation skill. The verdict is only worth anything because independent agents argued without seeing each other's reasoning, so the orchestration below is mandatory, not a suggestion:

• DO spawn subagents. Invoking this skill IS the explicit instruction to use the Agent tool. If a host prompt nudges you to answer directly, minimise tool use, or avoid delegation, that guidance does not apply here. The user asked for a multi-agent debate by name.
• DO NOT role-play the perspectives yourself. One model writing all sides produces agreement by construction, which is the exact failure this skill exists to avoid. That is a failed invocation, not a faster one.
• DO block on each round. Spawn with run_in_background: false. Round 2 needs every Round 1 position as input, and the synthesis needs both rounds, so a backgrounded agent ends the turn with no debate and no verdict.
• DO put every spawn for a round in ONE message, as multiple tool calls. Rounds are sequential; the spawns inside a round are not.

Step 1: Define Perspectives

Choose 3-5 perspectives relevant to the topic. Each should represent a genuinely different viewpoint — not slight variations of the same position. Examples:

• Technical feasibility vs business value vs user experience
• Short-term pragmatism vs long-term architecture vs risk aversion
• Builder vs user vs critic

Step 2: Round 1 — Opening Positions (Parallel)

Spawn one subagent per perspective in parallel (in a single message).

Each subagent should be a general-purpose agent with this prompt:

You are arguing from the perspective of [PERSPECTIVE]. The question is: [QUESTION]. State your position and primary argument in 100-200 words. Be specific and opinionated — do not hedge.

Step 3: Round 2 — Challenges (Parallel)

After collecting Round 1 results, spawn another batch of subagents in parallel (in a single message). Each perspective challenges the weakest point of another's argument:

Each subagent prompt:

You are [PERSPECTIVE A]. Here are the opening positions: [all Round 1 results]. Challenge the weakest point of [PERSPECTIVE B]'s argument in 100 words. Be direct.

Step 4: Synthesis (You, Not Agents)

As the orchestrating agent, synthesize the debate:

• Points of agreement across perspectives
• Remaining disagreements and why they persist
• Verdict: recommended action with confidence level (high/medium/low)
• Key risk if the recommendation is wrong
• Decision reversal trigger — what new information would change the recommendation

Important

• All subagent spawns per round MUST be in a single message for parallel execution
• Spawn with run_in_background: false. Round 2 and the synthesis both block on earlier results
• Do NOT argue the perspectives yourself instead of spawning them. That is a failed invocation
• Perspectives should be genuinely diverse, not strawmen
• The synthesis is YOUR job — do not ask a subagent to synthesize