save（save 搜索）

Save Session as Agent

Generate a reusable agent file from the current conversation and save it to .claude/agents/.

Instructions

Step 1: Generate the agent file

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:

1. User corrections are the most important signal. Every correction implies a rule the agent got wrong initially. Each correction MUST become an explicit rule.
2. Only capture what worked. If approach A failed and approach B worked — only document approach B.
3. Generalize — replace session-specific values (file names, URLs, credentials) with descriptive placeholders. The agent must work for similar tasks, not just this exact one.
4. 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>.

## Behavior

1. <First step the agent should take>
2. <Next step>
3. <...>

## Rules

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

## Output

<What the agent should produce — format, structure, location.>

Frontmatter Constraints

• 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)

Body Constraints

• 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

Guidelines

• 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

Step 2: Save the agent file

After generating the agent file content (starting with ---):

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

Step 3: Display the result

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.