creating-an-agent（创建 Agent）

Creating an Agent

REQUIRED BACKGROUND: Read ed3d-extending-claude:writing-claude-directives for foundational guidance on token efficiency, compliance techniques, and directive structure. This skill focuses on agent-specific patterns.

What is an Agent?

An agent is a specialized Claude instance with:

• Defined tools (Read, Edit, Bash, etc.)
• Specific responsibilities (code review, security audit, research)
• A focused system prompt

Agents are spawned via the Task tool or defined in plugin agents/ directories.

When to Create an Agent

Create when:

• Task requires specialized expertise
• Workflow benefits from tool restrictions
• You want consistent behavior across invocations
• Task is complex enough to warrant context isolation

Don't create for:

• Simple, one-off tasks
• Tasks the main Claude handles well
• Purely conversational interactions

Agent File Structure

agents/
  my-agent.md

Template:

---
name: agent-name
description: Use when [specific triggers] - [what agent does]
tools: Read, Grep, Glob, Bash
model: sonnet
---

# Agent Name

[Agent system prompt - who they are, what they do]

## Responsibilities
- Task 1
- Task 2

## Workflow
1. Step 1
2. Step 2

Description: The Critical Field

The description field determines when Claude auto-delegates to your agent. It's searched when matching tasks to available agents.

Writing Effective Descriptions

Format: "Use when [specific triggers/symptoms] - [what the agent does]"

Write in third person. Injected into system prompt.

# Bad: vague, no triggers
description: Helps with code

# Bad: first person
description: I review code for security issues

# Good: specific triggers + action
description: Use when reviewing code for security vulnerabilities, analyzing authentication flows, or checking for common security anti-patterns like SQL injection, XSS, or insecure dependencies

Include:

• Specific symptoms that trigger use
• Domain keywords (security, performance, testing)
• File types or patterns if relevant
• Actions the agent performs

Length: Max 1024 characters. Be specific, not verbose.

Tool Selection

Choose tools based on agent responsibilities:

Principle: Include only what the agent needs. Fewer tools = more focused behavior.

Example restrictions:

• Code reviewer: Read, Grep, Glob (no write access)
• Implementor: Read, Edit, Write, Bash, Grep, Glob
• Researcher: Read, WebFetch, WebSearch, Glob

Agent Prompt Structure

Role Definition

Start with who the agent is:

You are a security expert specializing in web application security and secure coding practices.

Responsibilities

Explicit, numbered list:

## Your Responsibilities

1. Identify security vulnerabilities
2. Review authentication logic
3. Check for insecure dependencies
4. Report findings with severity ratings

Workflow

Step-by-step process:

## Workflow

1. **Scan:** Use Grep to find common vulnerability patterns
2. **Analyze:** Use Read to examine flagged files
3. **Verify:** Use Bash to run security audit tools
4. **Report:** Provide structured findings

Output Format

Define expected structure:

## Reporting Format

For each finding:
- **Severity:** Critical/High/Medium/Low
- **Location:** `file:line`
- **Issue:** What's vulnerable
- **Impact:** What attacker could do
- **Fix:** How to remediate

Constraints

What the agent should NOT do:

## Constraints

- Report findings only; do not modify code
- Ask for clarification if scope is unclear
- Escalate to human for ambiguous security decisions

Model Selection

Specify in frontmatter:

model: opus

Testing Agents

1. Baseline Test

Run the task WITHOUT the agent. Document:

• What went wrong
• What was missing
• How long it took

2. Agent Test

Run with agent. Verify:

• Agent is auto-delegated (description triggers correctly)
• Workflow is followed
• Output matches expected format
• Tool restrictions are respected

3. Edge Case Testing

Test with:

• Ambiguous inputs
• Missing context
• Large/complex inputs
• Tasks outside scope (should refuse gracefully)

4. Iteration

If agent fails:

1. Identify root cause (description? workflow? constraints?)
2. Update agent definition
3. Re-test

Common Patterns

Code Reviewer

---
name: code-reviewer
description: Use when reviewing code changes, pull requests, or verifying implementation quality - analyzes for bugs, style issues, and best practices
tools: Read, Grep, Glob, Bash
model: opus
---

# Code Reviewer

You are a senior engineer reviewing code for correctness, readability, and maintainability.

## Responsibilities
1. Identify bugs and edge cases
2. Check error handling
3. Verify naming and style consistency
4. Suggest improvements

## Workflow
1. Read the changed files
2. Analyze for issues
3. Provide structured feedback

## Output Format
For each issue:
- **File:Line:** location
- **Severity:** Critical/Major/Minor
- **Issue:** description
- **Suggestion:** how to fix

Research Agent

---
name: researcher
description: Use when gathering information from the web, investigating APIs, or synthesizing documentation from multiple sources
tools: Read, WebFetch, WebSearch, Glob
model: sonnet
---

# Research Agent

You are a research specialist gathering and synthesizing information.

## Responsibilities
1. Search for relevant sources
2. Extract key information
3. Synthesize findings
4. Cite sources

## Workflow
1. WebSearch for relevant sources
2. WebFetch promising results
3. Extract and organize findings
4. Return structured synthesis with citations

Implementor Agent

---
name: task-implementor
description: Use when implementing specific tasks from plans - writes code, runs tests, commits changes following TDD workflow
tools: Read, Edit, Write, Bash, Grep, Glob, TaskCreate, TaskUpdate, TaskList
model: sonnet
---

# Task Implementor

You implement tasks following TDD principles.

## Responsibilities
1. Write failing test first
2. Implement minimal code to pass
3. Refactor if needed
4. Commit with descriptive message

## Constraints
- Never write implementation before test
- Run tests after each change
- Commit atomic, working changes only

Common Mistakes

Checklist

• Description starts with "Use when...", third person
• Description includes specific triggers/symptoms
• Tools restricted to necessary set
• Model appropriate for task complexity
• Responsibilities clearly listed
• Workflow is step-by-step
• Output format defined
• Constraints/limitations stated
• Tested for auto-delegation
• Tested for output quality