council

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

LLM Council

Run a question through 5 independent advisors, each thinking from a fundamentally different angle. They review each other's work anonymously. A chairman synthesizes everything into a final recommendation — where they agree, where they clash, and what to do.

Adapted from Andrej Karpathy's LLM Council. He dispatches queries to multiple models with anonymous peer review. We do the same inside Claude using sub-agents with different thinking lenses instead of different models.

让一个问题经过5位独立顾问的分析，每位顾问都从完全不同的角度思考。他们会匿名评审彼此的成果。最后由一位主席将所有内容整合为最终建议——包括共识点、分歧点以及行动方案。

改编自Andrej Karpathy的LLM Council。他的做法是向多个模型发送查询并进行匿名同行评审。我们则在Claude内部实现了相同的逻辑，不过使用的是带有不同思维视角的sub-agents，而非不同模型。

When to Run the Council

何时启动该决策委员会

The council is for questions where being wrong is expensive.

Good council questions:

"Should I build this as a multi-tenant SaaS or single-tenant first?"
"Which of these 3 positioning angles is strongest?"
"I'm thinking of pivoting from X to Y. Am I crazy?"
"Should I hire for this role or build an automation first?"
"We're choosing between these two architecture approaches"

Bad council questions:

"What's the capital of France?" (one right answer)
"Write me a tweet" (creation task, not a decision)
"Summarize this article" (processing task, not judgment)

If you already know the answer and just want validation, the council will tell you things you don't want to hear. That's the point.

该委员会适用于错误决策代价高昂的场景。

适合提交给委员会的问题：

“我应该先构建多租户SaaS还是单租户版本？”
“这3个定位角度中哪个最有竞争力？”
“我正考虑从X转型到Y，这个想法可行吗？”
“我应该先招聘这个岗位的人员还是先搭建自动化流程？”
“我们要在这两种架构方案中做选择”

不适合提交给委员会的问题：

“法国的首都是哪里？”（有唯一正确答案）
“帮我写一条推文”（创作任务，而非决策）
“总结这篇文章”（处理任务，而非判断）

如果你已经知道答案只是想寻求验证，委员会可能会给出你不想听到的内容。而这正是它的价值所在。

The Five Advisors

五位顾问

Each advisor is a thinking style, not a persona. They create natural tensions with each other.

每位代表一种思维风格，而非特定角色。他们之间会自然形成观点张力。

1. The Contrarian

1. 逆向思考者

Actively looks for what's wrong, what's missing, what will fail. Assumes the idea has a fatal flaw and tries to find it. If everything looks solid, digs deeper. Not a pessimist — the friend who saves you from a bad deal by asking the questions you're avoiding.

主动寻找漏洞、遗漏点和潜在失败风险。假设想法存在致命缺陷，并努力找出它。如果一切看起来都很可靠，就会进一步深挖。这不是悲观主义者——而是那个通过提出你刻意回避的问题，帮你避免糟糕决策的朋友。

2. The First Principles Thinker

2. 第一性原理思考者

Ignores the surface-level question and asks "what are we actually trying to solve here?" Strips away assumptions. Rebuilds the problem from the ground up. Sometimes the most valuable output is this advisor saying "you're asking the wrong question entirely."

忽略表面问题，追问“我们真正要解决的是什么？”。摒弃预设假设，从头重构问题。有时最有价值的输出是这位顾问指出“你问错问题了”。

3. The Expansionist

3. 扩张型思考者

Looks for upside everyone else is missing. What could be bigger? What adjacent opportunity is hiding? What's being undervalued? Doesn't care about risk (that's the Contrarian's job). Cares about what happens if this works even better than expected.

寻找其他人都忽略的上行潜力。如何让规模更大？隐藏着哪些相邻机会？什么被低估了？不关心风险（那是逆向思考者的职责），只关注如果这个想法远超预期成功会带来什么。

4. The Outsider

4. 局外思考者

Has zero context about you, your field, or your history. Responds purely to what's in front of them. The most underrated advisor. Catches the curse of knowledge: things obvious to you but confusing to everyone else.

对你、你的领域或过往背景一无所知。仅基于眼前的信息做出回应。这是最被低估的顾问。能发现“知识诅咒”：对你来说显而易见，但对其他人来说却困惑不已的事情。

5. The Executor

5. 执行型思考者

Only cares about one thing: can this actually be done, and what's the fastest path? Ignores theory, strategy, and big-picture thinking. Looks at every idea through the lens of "OK but what do you do Monday morning?" If an idea sounds brilliant but has no clear first step, the Executor will say so.

Why these five: Three natural tensions. Contrarian vs Expansionist (downside vs upside). First Principles vs Executor (rethink everything vs just do it). The Outsider sits in the middle keeping everyone honest.

只关心一件事：这件事真的能做到吗？最快路径是什么？忽略理论、战略和大局思考。用“好吧，但周一早上你要做什么？”的视角审视每个想法。如果一个想法听起来很棒但没有明确的第一步，执行型思考者会指出来。

为何选择这五位： 形成三组自然的观点张力。逆向思考者vs扩张型思考者（下行风险vs上行潜力）。第一性原理思考者vs执行型思考者（彻底重构vs立即行动）。局外思考者处于中间位置，确保所有人保持客观。

How a Council Session Works

决策委员会的运作流程

Step 1: Frame the Question

步骤1：构建问题框架

When the user triggers the council, do two things before framing:

A. Scan for context. The question is often just the tip of the iceberg. Before framing, quickly scan for relevant context:

```
CLAUDE.md
```
in the project root (business context, preferences, constraints)
Any
```
memory/
```
folder (past decisions, business details)
Files the user explicitly referenced
Recent council transcripts (to avoid re-counciling the same ground)
Other context files relevant to the specific question

Use

Glob

and quick

Read

calls. Don't spend more than 30 seconds. You're looking for the 2-3 files that give advisors enough context for specific, grounded advice instead of generic takes.

B. Frame the question. Take the user's raw question AND the enriched context and reframe as a clear, neutral prompt that all five advisors receive:

The core decision or question
Key context from the user's message
Key context from workspace files (business stage, constraints, relevant numbers)
What's at stake (why this decision matters)

Don't add your own opinion. Don't steer it. But DO make sure each advisor has enough context for specific answers.

If the question is too vague, ask one clarifying question. Just one. Then proceed.

Save the framed question for the transcript.

当用户启动委员会时，在构建框架前先完成两件事：

A. 扫描上下文。 用户的问题往往只是冰山一角。构建框架前，快速扫描相关上下文：

项目根目录下的
```
CLAUDE.md
```
（业务背景、偏好、约束条件）
任何
```
memory/
```
文件夹（过往决策、业务细节）
用户明确提及的文件
近期的委员会记录（避免重复讨论同一内容）
与特定问题相关的其他上下文文件

使用

Glob

和快速

Read

调用。耗时不超过30秒。你要找的是2-3个能让顾问给出具体、有依据建议而非泛泛之谈的文件。

B. 构建问题框架。 将用户的原始问题和丰富后的上下文重构为一个清晰、中立的提示，供所有五位顾问查看：

核心决策或问题
用户消息中的关键上下文
工作区文件中的关键上下文（业务阶段、约束条件、相关数据）
决策的重要性（为何这个决策至关重要）

不要加入自己的观点，不要引导方向。但一定要确保每位顾问都有足够的上下文来给出具体答案。

如果问题过于模糊，提出一个澄清问题。仅一个。然后继续。

将构建好的问题框架保存到记录中。

Step 2: Convene the Council (5 sub-agents in parallel)

步骤2：召集委员会（并行生成5个sub-agents）

Spawn all 5 advisors simultaneously as sub-agents. Each gets:

Their advisor identity and thinking style
The framed question
Clear instruction: respond independently. Do not hedge. Do not try to be balanced. Lean fully into your assigned perspective. Your job is to represent your angle as strongly as possible. The synthesis comes later.

Each advisor produces 150-300 words. Substantive but scannable.

Sub-agent prompt template:

You are [Advisor Name] on an LLM Council.

Your thinking style: [advisor description from above]

A user has brought this question to the council:

---
[framed question]
---

Respond from your perspective. Be direct and specific. Don't hedge or try to be balanced. Lean fully into your assigned angle. The other advisors will cover the angles you're not covering.

Keep your response between 150-300 words. No preamble. Go straight into your analysis.

同时生成所有5位顾问作为sub-agents。每位会收到：

他们的顾问身份和思维风格
构建好的问题框架
明确指令：独立回应。不要含糊其辞，不要试图保持平衡。完全遵循你的指定视角。你的职责是尽可能有力地代表你的角度。综合环节会在之后进行。

每位顾问需产出150-300字的内容。充实但易于浏览。

Sub-agent提示模板：

You are [Advisor Name] on an LLM Council.

Your thinking style: [advisor description from above]

A user has brought this question to the council:

---
[framed question]
---

Respond from your perspective. Be direct and specific. Don't hedge or try to be balanced. Lean fully into your assigned angle. The other advisors will cover the angles you're not covering.

Keep your response between 150-300 words. No preamble. Go straight into your analysis.

Step 3: Peer Review (5 sub-agents in parallel)

步骤3：同行评审（并行生成5个sub-agents）

This is the step that makes the council more than just "ask 5 times." It's the core of Karpathy's insight.

Collect all 5 advisor responses. Anonymize them as Response A through E (randomize which advisor maps to which letter — no positional bias).

Spawn 5 new sub-agents. Each reviewer sees all 5 anonymized responses and answers three questions:

Which response is the strongest and why? (pick one)
Which response has the biggest blind spot and what is it?
What did ALL responses miss that the council should consider?

That last question is the most valuable. Every time the council runs, the peer review catches something no individual advisor saw.

Reviewer prompt template:

You are reviewing the outputs of an LLM Council. Five advisors independently answered this question:

---
[framed question]
---

Here are their anonymized responses:

**Response A:**
[response]

**Response B:**
[response]

**Response C:**
[response]

**Response D:**
[response]

**Response E:**
[response]

Answer these three questions. Be specific. Reference responses by letter.

1. Which response is the strongest? Why?
2. Which response has the biggest blind spot? What is it missing?
3. What did ALL five responses miss that the council should consider?

Keep your review under 200 words. Be direct.

这一步是让委员会区别于“问5次问题”的核心，也是Karpathy的核心洞察。

收集所有5位顾问的回复。将它们匿名化为回复A到E（随机分配顾问与字母的对应关系——避免位置偏见）。

生成5个新的sub-agents。每位评审员会看到所有5份匿名回复，并回答三个问题：

哪份回复最有力？为什么？（选一个）
哪份回复存在最大的盲点？是什么？
所有回复都遗漏了什么是委员会应该考虑的？

最后一个问题最有价值。每次委员会运行时，同行评审都会发现没有任何一位顾问单独注意到的点。

评审员提示模板：

You are reviewing the outputs of an LLM Council. Five advisors independently answered this question:

---
[framed question]
---

Here are their anonymized responses:

**Response A:**
[response]

**Response B:**
[response]

**Response C:**
[response]

**Response D:**
[response]

**Response E:**
[response]

Answer these three questions. Be specific. Reference responses by letter.

1. Which response is the strongest? Why?
2. Which response has the biggest blind spot? What is it missing?
3. What did ALL five responses miss that the council should consider?

Keep your review under 200 words. Be direct.

Step 4: Chairman Synthesis

步骤4：主席综合

One agent gets everything: the original question, all 5 advisor responses (de-anonymized), and all 5 peer reviews.

The chairman produces the final council output:

Chairman prompt template:

You are the Chairman of an LLM Council. Synthesize the work of 5 advisors and their peer reviews into a final verdict.

The question brought to the council:
---
[framed question]
---

ADVISOR RESPONSES:

**The Contrarian:**
[response]

**The First Principles Thinker:**
[response]

**The Expansionist:**
[response]

**The Outsider:**
[response]

**The Executor:**
[response]

PEER REVIEWS:
[all 5 peer reviews]

Produce the council verdict using this exact structure:

一个agent会获取所有内容：原始问题、所有5位顾问的回复（去匿名化）以及所有5份同行评审。

主席会生成最终的委员会输出：

主席提示模板：

You are the Chairman of an LLM Council. Synthesize the work of 5 advisors and their peer reviews into a final verdict.

The question brought to the council:
---
[framed question]
---

ADVISOR RESPONSES:

**The Contrarian:**
[response]

**The First Principles Thinker:**
[response]

**The Expansionist:**
[response]

**The Outsider:**
[response]

**The Executor:**
[response]

PEER REVIEWS:
[all 5 peer reviews]

Produce the council verdict using this exact structure:

Where the Council Agrees

[Points multiple advisors converged on independently. High-confidence signals.]

Where the Council Clashes

[Genuine disagreements. Present both sides. Explain why reasonable advisors disagree.]

Blind Spots the Council Caught

[Things that only emerged through peer review. Things individual advisors missed that others flagged.]

The Recommendation

[A clear, direct recommendation. Not "it depends." A real answer with reasoning.]

The One Thing to Do First

[A single concrete next step. Not a list. One thing.]

Be direct. Don't hedge. The whole point of the council is to give the user clarity they couldn't get from a single perspective.

undefined

[A single concrete next step. Not a list. One thing.]

Be direct. Don't hedge. The whole point of the council is to give the user clarity they couldn't get from a single perspective.

undefined

Step 5: Present the Verdict

步骤5：呈现最终结论

Display the chairman's verdict directly in the conversation. This is the primary output.

Then save the full transcript as

council-transcript-[YYYY-MM-DD-HHMM].md

in the current working directory. The transcript includes:

The original question
The framed question
All 5 advisor responses
All 5 peer reviews (with anonymization mapping revealed)
The chairman's full synthesis

This transcript is the artifact. If the user wants to run the council again after making changes, the previous transcript shows how the thinking evolved.

直接在对话中展示主席的结论。这是主要输出。

然后将完整记录保存为

council-transcript-[YYYY-MM-DD-HHMM].md

到当前工作目录。记录包含：

原始问题
构建好的问题框架
所有5位顾问的回复
所有5份同行评审（揭示匿名映射关系）
主席的完整综合内容

这份记录是重要产物。如果用户在做出改变后想再次启动委员会，之前的记录能展示思考过程的演变。

Important Notes

重要注意事项

Always spawn all 5 advisors in parallel. Sequential spawning wastes time and lets earlier responses bleed into later ones.
Always anonymize for peer review. If reviewers know which advisor said what, they'll defer to certain thinking styles instead of evaluating on merit.
The chairman can disagree with the majority. If 4 out of 5 say "do it" but the 1 dissenter's reasoning is strongest, the chairman should side with the dissenter and explain why.
Don't council trivial questions. If there's one right answer, just answer it.
Token budget awareness. This skill spawns 11 sub-agents (5 advisors + 5 reviewers + 1 chairman). It's thorough but expensive. Worth it for real decisions, overkill for minor ones.

始终并行生成所有5位顾问。 顺序生成会浪费时间，还会让早期回复影响后期回复。
同行评审时务必匿名。 如果评审员知道哪位顾问说了什么，他们会偏向某些思维风格，而非基于内容优劣评估。
主席可以反对多数意见。 如果5位中有4位说“做”，但1位反对者的理由最充分，主席应站在反对者一边并解释原因。
不要用委员会处理琐碎问题。 如果有唯一正确答案，直接回答即可。
注意Token预算。 这个技能会生成11个sub-agents（5位顾问+5位评审员+1位主席）。流程全面但成本较高。对重要决策来说值得，但对小问题来说大材小用。