菜单
Use when a question has real trade-offs and no single right answer: architecture choices, library selection, design direction. Runs a multi-round debate where agents see each other's positions across rounds, converge, and a synthesis is produced. For factual questions, use consensus instead.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when a question has a factual or best-practice answer but a single response might hallucinate, or when the user asks to "get consensus" or "sanity check this with multiple agents". Spawns N subagents on the same question with varied personas and tallies answer frequency.
Use when triaging an inbox: fetching unprocessed emails, classifying each into a label taxonomy, and applying labels via the provider's API. Provider-agnostic skeleton; scripts/label.py must be wired to your email provider before first use.
Use when a workflow built on an MCP server has stabilized and should become a production skill, or when the user says "convert this MCP workflow to a skill". Produces a skill with a bundled script that calls the underlying API directly, eliminating the per-session MCP token overhead.
Use when a PRD or spec exists and needs to become an executable backlog. Breaks the document into GitHub issues structured as vertical-slice tracer bullets, marks each as HITL or AFK, and files a final QA issue. Requires the gh CLI.
Use when implementing any task that should end in a commit. Encodes the definition of done: understand, implement with red-green-refactor for backend logic, run typecheck and tests until clean, then commit. Also the unit of work for autonomous loops.
Use when creating or revising any Claude Code skill, or when the user says "write me a skill for X". Encodes the house conventions: frontmatter style, body structure, when to bundle scripts, and project vs user scope.
| name | debate |
| description | Use when a question has real trade-offs and no single right answer: architecture choices, library selection, design direction. Runs a multi-round debate where agents see each other's positions across rounds, converge, and a synthesis is produced. For factual questions, use consensus instead. |
Consensus counts independent answers. Debate makes the answers fight. Use it when the value is in the arguments, not the tally: trade-off decisions where you want the strongest case for each side before choosing.
Round 1 - positions. Spawn 3 subagents in parallel with the same fully-specified question and distinct stances: one argues the boldest option, one the most conservative, one optimizes for long-term maintenance. Each produces: position, top 3 arguments, and the strongest objection to itself it can think of.
Round 2 - rebuttals. Spawn 3 fresh subagents (subagents don't persist; the transcript is the memory). Each gets the question plus ALL Round 1 positions, assigned to defend one position: rebut the other two's best arguments, concede any point that is simply correct, update the position if a rebuttal fails. Forced concessions are the mechanism; a debate where nobody concedes anything is theater.
Round 3 - convergence (only if Round 2 moved positions). One subagent gets the full transcript and writes where the positions now agree, where they still differ, and what fact or constraint would settle each remaining difference.
You (the orchestrator) write the final answer, not a subagent:
Three agents, two to three rounds. More agents add cost faster than insight. If Round 1 comes back unanimous, stop there and say so; the debate was already settled.
