save

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Save Session as Agent

将会话保存为Agent

Generate a reusable agent file from the current conversation and save it to

.claude/agents/

从当前会话生成一个可复用的Agent文件，并保存到

.claude/agents/

目录下。

Instructions

操作说明

Step 1: Generate the agent file

步骤1：生成Agent文件

Analyze the entire conversation — the original task, every user correction, every tool call, and the final output — then distill it into a reusable agent file. The agent file is NOT a session log. It is a system prompt that a subagent will receive with no prior context.

Key priorities:

User corrections are the most important signal. Every correction implies a rule the agent got wrong initially. Each correction MUST become an explicit rule.
Only capture what worked. If approach A failed and approach B worked — only document approach B.
Generalize — replace session-specific values (file names, URLs, credentials) with descriptive placeholders. The agent must work for similar tasks, not just this exact one.
Keep it concise — this is a system prompt for a subagent. Shorter is better.

Output the agent file with YAML frontmatter followed by a system prompt body:

---
name: "<kebab-case-name>"
description: "<one-liner, max 200 chars>"
tools: Read, Glob, Grep, Bash, Write, Edit
model: sonnet
---

You are an agent that <role description — what this agent does>.

分析整个会话内容——包括初始任务、每一处用户修正、每一次工具调用以及最终输出——然后将其提炼为一个可复用的Agent文件。该Agent文件并非会话日志，而是一个供子Agent使用的无前置上下文的系统提示词。

Behavior

核心要点：

<First step the agent should take>
<Next step>
<...>

**用户修正是最重要的信号。**每一处修正都意味着Agent最初违反了某条规则，所有修正都必须转化为明确的规则。
**只保留有效的方法。**如果方法A失败而方法B有效，仅记录方法B。
进行泛化处理——将会话特有的值（如文件名、URL、凭证）替换为描述性占位符。该Agent需能处理同类任务，而非仅适用于当前这一项任务。
保持简洁——这是供子Agent使用的系统提示词，越简短越好。

输出的Agent文件需包含YAML前置元数据，后跟系统提示词主体：

---
name: "<kebab-case-name>"
description: "<one-liner, max 200 chars>"
tools: Read, Glob, Grep, Bash, Write, Edit
model: sonnet
---

You are an agent that <role description — what this agent does>.

Rules

Behavior

<Rule derived from user correction or session learning>
<Another rule>

<First step the agent should take>
<Next step>
<...>

Output

Rules

undefined

<Rule derived from user correction or session learning>
<Another rule>

Frontmatter Constraints

Output

```
name
```
— required, kebab-case, max 100 characters
```
description
```
— required, max 200 characters
```
tools
```
— required, comma-separated list of tools the agent needs (choose from: Read, Glob, Grep, Bash, Write, Edit, WebFetch, WebSearch)
```
model
```
— required, use
```
sonnet
```
unless the task clearly needs stronger reasoning (then use
```
opus
```
)

undefined

Body Constraints

前置元数据约束

Start with a one-sentence role description: "You are an agent that..."
Behavior section: numbered steps describing what the agent does, in order
Rules section: bullet list of constraints and guidelines — every user correction from the session MUST appear here
Output section: what the agent produces and in what format
All sections are required

```
name
```
— 必填项，采用kebab-case格式，最长100字符
```
description
```
— 必填项，最长200字符
```
tools
```
— 必填项，为Agent所需工具的逗号分隔列表（可选工具：Read, Glob, Grep, Bash, Write, Edit, WebFetch, WebSearch）
```
model
```
— 必填项，默认使用
```
sonnet
```
，除非任务明确需要更强的推理能力（此时使用
```
opus
```
）

Guidelines

主体内容约束

Write natural language instructions, not formal SHALL/MUST requirements
Be specific — "Use openpyxl for Excel files" not "Use the right tool"
Do NOT include session-specific details (specific file names, URLs, credentials, data values from this run)
DO generalize patterns — replace specific values with descriptive placeholders like
```
<input-file>
```
,
```
<target-url>
```
Only include steps that succeeded, not failed attempts
The agent file MUST be self-contained — the subagent needs nothing beyond this prompt and its input

以一句角色描述开头："You are an agent that..."
**Behavior（行为）**部分：按顺序列出Agent执行的步骤，采用编号格式
**Rules（规则）**部分：以项目符号列出约束和准则——会话中的每一处用户修正都必须包含在此
**Output（输出）**部分：说明Agent需要生成的内容及其格式
所有部分均为必填项

Step 2: Save the agent file

编写指南

After generating the agent file content (starting with

---

Extract the
```
name
```
from the YAML frontmatter. Use it as the slug directly (it's already kebab-case).
Create the
```
.claude/agents/
```
directory if it doesn't exist:
bash
```
mkdir -p .claude/agents
```
Write the agent file content to
```
.claude/agents/{name}.md
```
using the Write tool.

使用自然语言编写指令，而非正式的SHALL/MUST类强制表述
表述要具体——比如写"Use openpyxl for Excel files"而非"Use the right tool"
请勿包含会话特有的细节（如本次会话中的具体文件名、URL、凭证、数据值）
务必对模式进行泛化——将特定值替换为描述性占位符，如
```
<input-file>
```
、
```
<target-url>
```
仅保留成功的步骤，不要包含失败的尝试
Agent文件必须是自包含的——子Agent仅需该提示词和输入即可运行

Step 3: Display the result

步骤2：保存Agent文件

Tell the user the agent was saved to

.claude/agents/{name}.md

. Let them know they can invoke it with

@{name}

in any conversation. Since the agent lives in the repo, it's automatically shared with anyone who has access to the repository.

生成Agent文件内容（以

---

开头）后：

从YAML前置元数据中提取
```
name
```
字段，直接将其作为文件名（已为kebab-case格式）。
如果
```
.claude/agents/
```
目录不存在则创建：
bash
```
mkdir -p .claude/agents
```
使用Write工具将Agent文件内容写入
```
.claude/agents/{name}.md
```
。

—

步骤3：展示结果

—

告知用户Agent已保存至

.claude/agents/{name}.md

。同时说明他们可在任意会话中通过

@{name}

调用该Agent。由于Agent存储在代码仓库中，所有拥有仓库访问权限的用户都能自动获取该Agent。