agent-development（Agent 开发）

Agent Development for Claude Code

Build effective custom agents for Claude Code with proper delegation, tool access, and prompt design.

Agent Description Pattern

The description field determines whether Claude will automatically delegate tasks.

Strong Trigger Pattern

---
name: agent-name
description: |
  [Role] specialist. MUST BE USED when [specific triggers].
  Use PROACTIVELY for [task category].
  Keywords: [trigger words]
tools: Read, Write, Edit, Glob, Grep, Bash
model: sonnet
---

Weak vs Strong Descriptions

Key phrases:

• "MUST BE USED when..."
• "Use PROACTIVELY for..."
• Include trigger keywords

Delegation Mechanisms

1. Explicit: Task tool subagent_type: "agent-name" - always works
2. Automatic: Claude matches task to agent description - requires strong phrasing

Session restart required after creating/modifying agents.

Tool Access Principle

If an agent doesn't need Bash, don't give it Bash.

Why? Models default to cat > file << 'EOF' heredocs instead of Write tool. Each bash command requires approval, causing dozens of prompts per agent run.

Allowlist Pattern

Instead of restricting Bash, allowlist safe commands in .claude/settings.json:

{
  "permissions": {
    "allow": [
      "Write", "Edit", "WebFetch(domain:*)",
      "Bash(cd *)", "Bash(cp *)", "Bash(mkdir *)", "Bash(ls *)",
      "Bash(cat *)", "Bash(head *)", "Bash(tail *)", "Bash(grep *)",
      "Bash(diff *)", "Bash(mv *)", "Bash(touch *)", "Bash(file *)"
    ]
  }
}

Model Selection (Quality First)

Don't downgrade quality to work around issues - fix root causes instead.

Memory Limits

Root Cause Fix (REQUIRED)

Add to ~/.bashrc or ~/.zshrc:

export NODE_OPTIONS="--max-old-space-size=16384"

Increases Node.js heap from 4GB to 16GB.

Parallel Limits (Even With Fix)

Recovery

1. source ~/.bashrc or restart terminal
2. NODE_OPTIONS="--max-old-space-size=16384" claude
3. Check what files exist, continue from there

Sub-Agent vs Remote API

Always prefer Task sub-agents over remote API calls.

// ❌ WRONG - Remote API call
const response = await fetch('https://api.anthropic.com/v1/messages', {...})

// ✅ CORRECT - Use Task tool
// Invoke Task with subagent_type: "general-purpose"

Declarative Over Imperative

Describe what to accomplish, not how to use tools.

Wrong (Imperative)

### Check for placeholders

grep -r "PLACEHOLDER:" build/*.html

Right (Declarative)

### Check for placeholders
Search all HTML files in build/ for:
- PLACEHOLDER: comments
- TODO or TBD markers
- Template brackets like [Client Name]

Any match = incomplete content.

What to Include

Self-Documentation Principle

"Agents that won't have your context must be able to reproduce the behaviour independently."

Every improvement must be encoded into the agent's prompt, not left as implicit knowledge.

What to Encode

Test: Would a Fresh Agent Succeed?

Before completing any agent improvement:

1. Read the agent prompt as if you have no context
2. Ask: Could a new session follow this and produce the same quality?
3. If no: Add missing instructions, patterns, or references

Anti-Patterns

Flexibility vs Rigidity

Match specification level to task type. Over-specifying flexible agents makes them brittle.

Signs You've Over-Specified

• Agent fills in template sections with "N/A"
• Agent tries to complete all phases even when irrelevant
• Scoring systems produce meaningless numbers
• Agent fails when scope doesn't match assumptions
• Long agents (>150 lines) for simple tasks

Flexible Agent Guidelines

DO:

• Describe *what to look for*, not exact steps
• Provide output *examples*, not rigid templates
• Include scope control ("if >30 items, ask user")
• Give escape hatches ("if unsure, flag for review")
• Keep under 100 lines for judgment tasks

DON'T:

• Require filling every section of a template
• Create elaborate weighted scoring systems
• List every possible check exhaustively
• Assume scope without asking

Example: Docs Auditor

Over-specified (bad):

## Phase 1: Discovery
Execute Glob for all .md files...

## Phase 6: Generate Report
| Category | Weight | Score | Weighted |
|----------|--------|-------|----------|
| Links    | 20%    | X/100 | X        |

Right-sized (good):

## What to Check
- TODOs, broken links, stale versions

## Output Format
List issues by severity. Include file:line and fix.

## Scope Control
If >30 files, ask user which to focus on.

Agent Prompt Structure

Effective agent prompts include:

## Your Role
[What the agent does]

## Blocking Check
[Prerequisites that must exist]

## Input
[What files to read]

## Process
[Step-by-step with encoded learnings]

## Output
[Exact file paths and formats]

## Quality Checklist
[Verification steps including learned gotchas]

## Common Issues
[Patterns discovered during development]

Pipeline Agents

When inserting a new agent into a numbered pipeline (e.g., HTML-01 → HTML-05 → HTML-11):

Common bug: New agent is "orphaned" because predecessor still points to old next agent.

Verification:

grep -n "Next:.*→\|Then.*runs next" .claude/agents/*.md

The Sweet Spot

Best use case: Tasks that are repetitive but require judgment.

Example: Auditing 70 skills manually = tedious. But each audit needs intelligence (check docs, compare versions, decide what to fix). Perfect for parallel agents with clear instructions.

Not good for:

• Simple tasks (just do them)
• Highly creative tasks (need human direction)
• Tasks requiring cross-file coordination (agents work independently)

Effective Prompt Template

For each [item]:
1. Read [source file]
2. Verify with [external check - npm view, API call, etc.]
3. Check [authoritative source]
4. Score/evaluate
5. FIX issues found ← Critical instruction

Key elements:

• "FIX issues found" - Without this, agents only report. With it, they take action.
• Exact file paths - Prevents ambiguity
• Output format template - Ensures consistent, parseable reports
• Batch size ~5 items - Enough work to be efficient, not so much that failures cascade

Workflow Pattern

1. ME: Launch 2-3 parallel agents with identical prompt, different item lists
2. AGENTS: Work in parallel (read → verify → check → edit → report)
3. AGENTS: Return structured reports (score, status, fixes applied, files modified)
4. ME: Review changes (git status, spot-check diffs)
5. ME: Commit in batches with meaningful changelog
6. ME: Push and update progress tracking

Why agents don't commit: Allows human review, batching, and clean commit history.

Signs a Task Fits This Pattern

Good fit:

• Same steps repeated for many items
• Each item requires judgment (not just transformation)
• Items are independent (no cross-item dependencies)
• Clear success criteria (score, pass/fail, etc.)
• Authoritative source exists to verify against

Bad fit:

• Items depend on each other's results
• Requires creative/subjective decisions
• Single complex task (use regular agent instead)
• Needs human input mid-process

Quick Reference

Agent Frontmatter Template

---
name: my-agent
description: |
  [Role] specialist. MUST BE USED when [triggers].
  Use PROACTIVELY for [task category].
  Keywords: [trigger words]
tools: Read, Write, Edit, Glob, Grep, Bash
model: sonnet
---

Fix Bash Approval Spam

1. Remove Bash from tools if not needed
2. Put critical instructions FIRST (right after frontmatter)
3. Use allowlists in .claude/settings.json

Memory Crash Recovery

export NODE_OPTIONS="--max-old-space-size=16384"
source ~/.bashrc && claude