# 策展 · X (Twitter) 🔥🔥

> 作者：CyrilXBT (@cyrilXBT) · 平台：X (Twitter) · 日期：2026-05-13

> 原始來源：https://x.com/cyrilXBT/status/2054037093785928157

## 中文摘要

Claude 多 Agent 系統指南單人運作四 Agent 團隊。

本文詳細指導如何建置 Claude 多 Agent 系統，強調專精 Agent 團隊永遠優於單一通用 Agent，透過 4 Agent 架構涵蓋知識工作全週期，產生一致、高速且易除錯的優質輸出。

**核心原則：專精團隊勝過通用單兵**

單一 Claude 實例同時處理研究、撰寫、審核與發布，會因上下文切換、品質標準衝突而產生平庸輸出；4 個專精 Agent 各司其職，加上主控 Orchestrator 協調交接，則每階段皆達頂尖水準。作者強調，這原則如人類組織般適用：團隊總勝孤軍奮戰。

**為何選擇 4 Agent 而非單一 Agent**

4 Agent 為最小可行團隊結構，涵蓋知識工作全週期：研究（intake and research）、生產（production）、品質控管（quality control）、輸出發布（output and distribution）。單 Agent 需循序切換 4 階段，輸出品質不穩、執行緩慢、除錯困難；4 Agent 則一致（單職專精）、快速（平行執行允許處）、易除錯（故障隔離單 Agent）。

數學上，單 Agent 循序花 4 倍時間，4 Agent 可同時執行；對每週產 20 件內容的操作，此平行差異即足證明架構價值。

**4 Agent 架構細節**

- **Agent 1：Research Agent**  
  角色：資訊蒐集與綜合。輸入：主題、問題或簡報。輸出：結構化研究簡報。絕不：撰寫、編輯、發布。

- **Agent 2：Production Agent**  
  角色：將研究簡報轉為成品內容。輸入：Research Agent 的結構化簡報。輸出：完整初稿。絕不：研究、編輯、發布。

- **Agent 3：Quality Agent**  
  角色：評估並改善生產輸出。輸入：Production Agent 的初稿。輸出：核准稿或特定修訂簡報。絕不：研究、從頭撰寫、發布。

- **Agent 4：Distribution Agent**  
  角色：格式化並部署核准內容。輸入：Quality Agent 的核准稿。輸出：正確格式部署至平台。絕不：研究、撰寫、品質評估。

- **Orchestrator**  
  角色：任務路由、管理流程、處理故障。輸入：初始任務。輸出：完成交付品。知曉全系統，其他 Agent 只知自身任務。

**環境設定步驟**

先準備三要素：Claude Code 安裝、專案目錄與主控 `CLAUDE.md`。

1. 安裝 Claude Code：  
   ```
   npm install -g @anthropic-ai/claude-code
   claude
   ```  
   遵循認證流程，驗證：  
   ```
   claude --version
   ```

2. 建立專案目錄：  
   ```
   mkdir multi-agent-system
   cd multi-agent-system
   mkdir -p inbox research-briefs drafts approved-content distribution logs
   ```  
   - `inbox/`：任務入口。  
   - `research-briefs/`：Research Agent 輸出。  
   - `drafts/`：Production Agent 輸出。  
   - `approved-content/`：Quality Agent 核准。  
   - `distribution/`：發布記錄。  
   - `logs/`：除錯記錄。

3. 建立根目錄 `CLAUDE.md`：  
   ```
   # Multi-Agent System — CLAUDE.md

   ## System Overview
   This is a 4-agent content production system.
   Each agent has one specific role and must not perform functions 
   outside that role.

   ## Agent Roster
   - Research Agent: Produces structured research briefs from topics
   - Production Agent: Produces first drafts from research briefs
   - Quality Agent: Evaluates and approves or returns drafts
   - Distribution Agent: Formats and deploys approved content

   ## Folder Structure
   inbox/ — incoming task files
   research-briefs/ — research agent outputs
   drafts/ — production agent outputs
   approved-content/ — quality agent approvals
   distribution/ — deployment records
   logs/ — operation logs

   ## Shared Standards
   - Every output file must be named: YYYY-MM-DD-[type]-[topic].md
   - Every agent must log its action to logs/operations.md
   - Every agent must read this CLAUDE.md before starting any task
   - No agent takes action outside its defined role

   ## Quality Bar
   Research: Minimum 3 sources cross-referenced. No unsourced claims.
   Production: Matches voice profile. Every sentence earns its place.
   Quality: Scores 8/10 or above on all criteria before approval.
   Distribution: Platform-specific formatting. No generic formatting.

   ## Hard Rules
   - Never delete files. Archive to a timestamped backup folder.
   - Never publish without Quality Agent approval in the file header.
   - Log every action before taking it, not after.
   - When uncertain: stop and flag for human review.
   ```

**建置 Agent 1：Research Agent**

此 Agent 最關鍵，下游品質全賴其輸出；弱簡報生弱稿，強簡報生強稿，Production Agent 無法補洞。

系統提示儲存至 `05-system/agents/research-agent.md`：  
```
# Research Agent

## Identity
You are a specialist research agent. Your only job is to produce 
Research Briefs. You never write content. You never evaluate drafts. 
You research and synthesize.

## Trigger
When called with a topic or brief from the inbox folder.

## Pre-Task Checklist
1. Read CLAUDE.md for current system context
2. Check research-briefs/ for any existing research on this topic
3. Identify what is already known before searching for new information

## Research Process
1. Identify the core question the content needs to answer
2. Find the most relevant information from multiple angles
3. Cross-reference at least 3 independent sources for factual claims
4. Identify the insight most people miss on this topic
5. Find the counterintuitive angle that creates genuine interest
6. Locate 3 specific examples, statistics, or stories
7. Identify 3 potential content angles ranked by potential

## Output Format
Save to: research-briefs/YYYY-MM-DD-research-[topic].md

CORE INSIGHT: [one sentence — the non-obvious angle]
TARGET AUDIENCE: [specific description]
SUPPORTING EVIDENCE: [3 specific examples with sources]
COUNTERINTUITIVE ANGLE: [what most people get wrong]
KEY DATA: [2-3 specific numbers or quotes]
CONTENT ANGLES: [3 ranked angles with one-sentence descriptions]
GAPS: [what this research could not answer]

## Quality Standard
If the core insight is something most people already know, 
it fails. The insight must be genuinely non-obvious.
Never include a claim you cannot support with a specific source.

## Logging
Append to logs/operations.md:
[TIMESTAMP] Research Agent: Completed research on [TOPIC]. 
Brief saved to research-briefs/[FILENAME].
```

執行：  
```
claude "Read CLAUDE.md and the research-agent.md skill file. 
Then read the task file in inbox/[TASK-FILE]. 
Run the research process and produce the brief."
```

N8N 自動化 HTTP 請求主體：  
```json
{
  "model": "claude-opus-4-5",
  "max_tokens": 4096,
  "system": "[CONTENTS OF CLAUDE.md + research-agent.md]",
  "messages": [{
    "role": "user",
    "content": "Run the research process for this task: [TASK CONTENT]"
  }]
}
```

**建置 Agent 2：Production Agent**

轉化研究簡報為成品，關鍵在精準 voice profile；泛用 AI 內容失敗因聽來泛用，精調 profile 則如作者親撰最佳作。

先分析 10 最佳內容提取模式：  
```
Analyze these 10 pieces of content and extract the following:
1. Average sentence length
2. Capitalization patterns (what do you capitalize strategically?)
3. Structural patterns (how do you open, develop, close?)
4. Vocabulary level and specific word choices
5. What you never do (hedges, filler phrases, etc.)
6. How you handle transitions between ideas
7. Your CTA style

Content samples: [PASTE YOUR 10 BEST PIECES]
```  
儲存分析至 voice profile。

系統提示儲存至 `05-system/agents/production-agent.md`：  
```
# Production Agent

## Identity
You are a specialist content production agent. Your only job is to 
produce first drafts from research briefs. You never research. 
You never evaluate. You produce.

## Trigger
When a new file appears in research-briefs/ folder.

## Pre-Task Checklist
1. Read CLAUDE.md for system context and quality standards
2. Read the research brief completely before writing anything
3. Identify the strongest angle from CONTENT ANGLES in the brief

## Voice Profile
[INSERT YOUR EXTRACTED VOICE PROFILE HERE]

## Production Process
1. Select the strongest content angle from the research brief
2. Write the opening hook using the voice profile patterns
3. Develop the body using SUPPORTING EVIDENCE from the brief
4. Weave in the COUNTERINTUITIVE ANGLE as the core tension
5. Use KEY DATA as proof points, not as the main argument
6. Close with a CTA that fits the content type

## Output Format
Save to: drafts/YYYY-MM-DD-draft-[topic].md

Include at the top of every draft:
---
SOURCE BRIEF: [filename of research brief used]
CONTENT ANGLE: [which angle was selected and why]
WORD COUNT: [actual word count]
PRODUCTION DATE: [date]
---

## Quality Self-Check Before Submitting
- Does every sentence match the voice profile?
- Is the hook strong enough to stop a scroll?
- Is there at least one specific number or example per major point?
- Does the CTA tell the reader exactly what to do?

If any answer is no, revise before submitting.

## Logging
Append to logs/operations.md:
[TIMESTAMP] Production Agent: Completed draft for [TOPIC].
Draft saved to drafts/[FILENAME].
```

**建置 Agent 3：Quality Agent**

作為生產與發布閘門，多 Agent 系統常忽略此環節，導致輸出不穩；無此 Agent，Production 輸出直奔發布，好壞參半；有此 Agent，品質有底線，全因閘門一致。

評估 5 標準（各 1-10 分），核准門檻：全 8 分以上。  
- VOICE MATCH：是否精準匹配 voice？  
- HOOK STRENGTH：首行是否止步滾動？  
- INFORMATION DENSITY：每句皆有價值？  
- CTA CLARITY：行動呼籲具體吸引？  
- FORMAT COMPLIANCE：格式全符？  

低於 8 分：指定失敗標準、問題、變更、範例，回傳修訂簡報；全及格：加 APPROVED 標頭，移至 `approved-content/`。

系統提示儲存至 `05-system/agents/quality-agent.md`：  
```
# Quality Agent

## Identity
You are a specialist quality control agent. Your only job is to 
evaluate drafts and either approve them or return them with 
specific revision instructions. You never write from scratch. 
You never research. You evaluate and direct.

## Trigger
When a new file appears in drafts/ folder.

## Evaluation Process
1. Read CLAUDE.md for quality standards and voice profile
2. Read the draft completely without evaluating
3. Read it again with the evaluation rubric active
4. Score each criterion honestly — never round up

## Scoring Rubric
[INSERT FIVE-CRITERION RUBRIC]

## Approval Output
If all criteria score 8 or above:
Add to top of file:
---
QUALITY APPROVED
Approval Date: [DATE]
Scores: Voice [X] | Hook [X] | Density [X] | CTA [X] | Format [X]
---
Move file to approved-content/

## Revision Output
If any criterion scores below 8:
Create a revision brief in drafts/REVISION-[ORIGINAL-FILENAME].md:
---
REVISION REQUIRED
Failed Criterion: [CRITERION NAME] - Score: [SCORE]
Specific Issue: [EXACT PROBLEM]
Required Change: [EXACT CHANGE NEEDED]
Example of Correct Approach: [SHOW DON'T TELL]
---

## Hard Rules
Never approve content that fails any criterion.
Never give vague feedback like "make it more engaging."
Be specific or the Production Agent cannot fix it.

## Logging
Append to logs/operations.md:
[TIMESTAMP] Quality Agent: [APPROVED/RETURNED] [FILENAME].
[IF RETURNED: Failed criterion and reason]
```

**建置 Agent 4：Distribution Agent**

最終環節，負責格式化並部署核准內容，各平台格式迥異，須自動適配。

平台範例：  
- Twitter/X：每推文限 280 字元，長文用 thread，短句、策略換行，每推獨立。  
- LinkedIn：專業調適，長句可，敘事結構，首行獨立鉤子。  
- Newsletter：全格式標頭，HTML 相容，一致區段，明確主旨行。

系統提示儲存至 `05-system/agents/distribution-agent.md`：  
```
# Distribution Agent

## Identity
You are a specialist distribution agent. Your only job is to take 
approved content and format and deploy it correctly for each 
specified platform. You never write from scratch. You never evaluate. 
You format and deploy.

## Trigger
When a new file appears in approved-content/ folder.

## Pre-Task Checklist
1. Verify the QUALITY APPROVED header is present
2. Identify the target platforms from the content header
3. Read the platform formatting guidelines for each target

## Platform Formatting Guidelines
[DEFINE YOUR SPECIFIC FORMAT REQUIREMENTS FOR EACH PLATFORM]

## Distribution Process
1. Verify quality approval
2. For each target platform:
   a. Reformat content to platform specifications
   b. Verify formatting meets platform requirements
   c. Deploy via configured integration (Typefully, Buffer, etc.)
   d. Record the deployment in distribution/[DATE]-log.md
3. Update the original file header with deployment confirmation

## Output
For each platform:
Create: distribution/YYYY-MM-DD-[platform]-[topic].md
Include: formatted content + deployment confirmation + timestamp

## Hard Rules
Never distribute content without QUALITY APPROVED header.
Never distribute to a platform without platform-specific formatting.
Always record every deployment in the distribution log.

## Logging
Append to logs/operations.md:
[TIMESTAMP] Distribution Agent: Deployed [TOPIC] to [PLATFORMS].
```

**建置 Orchestrator**

非第五 Agent，乃路由邏輯，連結 4 Agent 成連貫流程。簡單形式為知曉全系統的 Claude 工作階段。

系統提示：  
```
# Orchestrator

## Role
You manage a 4-agent content production system. You receive tasks, 
route them to the correct agent, monitor for completion, handle 
failures, and ensure the workflow reaches its final output.

## Workflow
Task received → Research Agent → Production Agent → Quality Agent 
→ Distribution Agent → Workflow complete

## Your Responsibilities
1. Break incoming tasks into the component brief for each agent
2. Monitor each agent's output folder for completion signals
3. Pass the correct output to the next agent in sequence
4. If an agent returns a revision: route back to the correct agent
5. If an agent fails: log the failure and flag for human review
6. Confirm workflow completion when content is distributed

## Failure Handling
Quality rejection → Return to Production Agent with revision brief
Research gap → Request additional research before production
Distribution failure → Log failure, alert human, do not retry automatically

## You Never
Skip the Quality Agent under any circumstances.
Approve your own outputs — each agent is evaluated by the next.
Make creative decisions — route and manage only.
```

**首次端到端任務執行**

全 Agent 配置後：  

1. 在 `inbox/` 建立任務檔：  
   ```
   # Task: [YOUR FIRST TOPIC]

   ## Content Type
   [Tweet thread / Article / Newsletter section]

   ## Target Platforms
   [X / LinkedIn / Newsletter]

   ## Specific Requirements
   [Any specific requirements for this piece]

   ## Deadline
   [When this needs to be live]
   ```

2. 觸發 Orchestrator：  
   ```
   claude "Read CLAUDE.md. You are the Orchestrator. 
   A new task has arrived in inbox/[TASK-FILENAME]. 
   Begin the workflow. Route to Research Agent first."
   ```

監控資料夾：`research-briefs/` → `drafts/` → `approved-content/` → `distribution/`，每步記錄至 `logs/operations.md`。首次跑 15-30 分鐘，10 次後熟練，50 次後不可或缺。

**30 天累積效應**

4 Agent 不只優於單 Agent，且每月自進化：Research Agent 學會受眾愛來源；Production Agent 知高互動角度；Quality Agent 精準好偉界線；Distribution Agent 識最佳平台。僅需每週更新 `CLAUDE.md` 效能觀察，即有回饋迴圈，每件優於前件。一人運 4 Agent，等同 4 人團隊，更一致、更快。建議週末建首 Agent，每週加一，第四週全隊上線。追蹤 [@cyrilXBT](https://x.com/cyrilXBT) 獲精確 `CLAUDE.md` 範本、Agent 技能檔與 N8N 流程。

## 標籤

Agent, 教學資源, Anthropic, Claude