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

Weak (won't auto-delegate)	Strong (auto-delegates)
"Analyzes screenshots for issues"	"Visual QA specialist. MUST BE USED when analyzing screenshots. Use PROACTIVELY for visual QA."
"Runs Playwright scripts"	"Playwright specialist. MUST BE USED when running Playwright scripts. Use PROACTIVELY for browser automation."

Key phrases:

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

Delegation Mechanisms

Explicit:
```
Task tool subagent_type: "agent-name"
```
- always works
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.

Agent needs to...	Give tools	Don't give
Create files only	Read, Write, Edit, Glob, Grep	Bash
Run scripts/CLIs	Read, Write, Edit, Glob, Grep, Bash	—
Read/audit only	Read, Glob, Grep	Write, Edit, 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.

Model	Use For
Opus	Creative work (page building, design, content) - quality matters
Sonnet	Most agents - content, code, research (default)
Haiku	Only script runners where quality doesn't matter

Memory Limits

Root Cause Fix (REQUIRED)

Add to

~/.bashrc

~/.zshrc

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

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

Parallel Limits (Even With Fix)

Agent Type	Max Parallel	Notes
Any agents	2-3	Context accumulates; batch then pause
Heavy creative (Opus)	1-2	Uses more memory

Recovery

```
source ~/.bashrc
```
or restart terminal

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

Check what files exist, continue from there

Sub-Agent vs Remote API

Always prefer Task sub-agents over remote API calls.

Aspect	Remote API Call	Task Sub-Agent
Tool access	None	Full (Read, Grep, Write, Bash)
File reading	Must pass all content in prompt	Can read files iteratively
Cross-referencing	Single context window	Can reason across documents
Decision quality	Generic suggestions	Specific decisions with rationale
Output quality	~100 lines typical	600+ lines with specifics

// ❌ 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
```bash
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.
</code></pre>
<h3>What to Include</h3>
<table><thead><tr><th>Include</th><th>Skip</th></tr></thead><tbody><tr><td>Task goal and context</td><td>Explicit bash/tool commands</td></tr><tr><td>Input file paths</td><td>&quot;Use X tool to...&quot;</td></tr><tr><td>Output file paths and format</td><td>Step-by-step tool invocations</td></tr><tr><td>Success/failure criteria</td><td>Shell pipeline syntax</td></tr><tr><td>Blocking checks (prerequisites)</td><td>Micromanaged workflows</td></tr><tr><td>Quality checklists</td><td></td></tr></tbody></table>
<h2>Self-Documentation Principle</h2>
<blockquote>
<p>&quot;Agents that won&#x27;t have your context must be able to reproduce the behaviour independently.&quot;</p>
</blockquote>
<p>Every improvement must be encoded into the agent&#x27;s prompt, not left as implicit knowledge.</p>
<h3>What to Encode</h3>
<table><thead><tr><th>Discovery</th><th>Where to Capture</th></tr></thead><tbody><tr><td>Bug fix pattern</td><td>Agent&#x27;s &quot;Corrections&quot; or &quot;Common Issues&quot; section</td></tr><tr><td>Quality requirement</td><td>Agent&#x27;s &quot;Quality Checklist&quot; section</td></tr><tr><td>File path convention</td><td>Agent&#x27;s &quot;Output&quot; section</td></tr><tr><td>Tool usage pattern</td><td>Agent&#x27;s &quot;Process&quot; section</td></tr><tr><td>Blocking prerequisite</td><td>Agent&#x27;s &quot;Blocking Check&quot; section</td></tr></tbody></table>
<h3>Test: Would a Fresh Agent Succeed?</h3>
<p>Before completing any agent improvement:</p>
<ol>
<li>Read the agent prompt as if you have no context</li>
<li>Ask: Could a new session follow this and produce the same quality?</li>
<li>If no: Add missing instructions, patterns, or references</li>
</ol>
<h3>Anti-Patterns</h3>
<table><thead><tr><th>Anti-Pattern</th><th>Why It Fails</th></tr></thead><tbody><tr><td>&quot;As we discussed earlier...&quot;</td><td>No prior context exists</td></tr><tr><td>Relying on files read during dev</td><td>Agent may not read same files</td></tr><tr><td>Assuming knowledge from errors</td><td>Agent won&#x27;t see your debugging</td></tr><tr><td>&quot;Just like the home page&quot;</td><td>Agent hasn&#x27;t built home page</td></tr></tbody></table>
<h2>Agent Prompt Structure</h2>
<p>Effective agent prompts include:</p>
<pre><code class="language-markdown">## 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]
</code></pre>
<h2>Pipeline Agents</h2>
<p>When inserting a new agent into a numbered pipeline (e.g., <code>HTML-01</code> → <code>HTML-05</code> → <code>HTML-11</code>):</p>
<table><thead><tr><th>Must Update</th><th>What</th></tr></thead><tbody><tr><td>New agent</td><td>&quot;Workflow Position&quot; diagram + &quot;Next&quot; field</td></tr><tr><td><strong>Predecessor agent</strong></td><td>Its &quot;Next&quot; field to point to new agent</td></tr></tbody></table>
<p><strong>Common bug</strong>: New agent is &quot;orphaned&quot; because predecessor still points to old next agent.</p>
<p><strong>Verification</strong>:</p>
<pre><code class="language-bash">grep -n &quot;Next:.*→\|Then.*runs next&quot; .claude/agents/*.md
</code></pre>
<h2>The Sweet Spot</h2>
<p><strong>Best use case</strong>: Tasks that are <strong>repetitive but require judgment</strong>.</p>
<p>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.</p>
<p><strong>Not good for</strong>:</p>
<ul>
<li>Simple tasks (just do them)</li>
<li>Highly creative tasks (need human direction)</li>
<li>Tasks requiring cross-file coordination (agents work independently)</li>
</ul>
<h2>Effective Prompt Template</h2>
<pre><code>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
</code></pre>
<p><strong>Key elements</strong>:</p>
<ul>
<li><strong>&quot;FIX issues found&quot;</strong> - Without this, agents only report. With it, they take action.</li>
<li><strong>Exact file paths</strong> - Prevents ambiguity</li>
<li><strong>Output format template</strong> - Ensures consistent, parseable reports</li>
<li><strong>Batch size ~5 items</strong> - Enough work to be efficient, not so much that failures cascade</li>
</ul>
<h2>Workflow Pattern</h2>
<pre><code>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
</code></pre>
<p><strong>Why agents don&#x27;t commit</strong>: Allows human review, batching, and clean commit history.</p>
<h2>Signs a Task Fits This Pattern</h2>
<p><strong>Good fit</strong>:</p>
<ul>
<li>Same steps repeated for many items</li>
<li>Each item requires judgment (not just transformation)</li>
<li>Items are independent (no cross-item dependencies)</li>
<li>Clear success criteria (score, pass/fail, etc.)</li>
<li>Authoritative source exists to verify against</li>
</ul>
<p><strong>Bad fit</strong>:</p>
<ul>
<li>Items depend on each other&#x27;s results</li>
<li>Requires creative/subjective decisions</li>
<li>Single complex task (use regular agent instead)</li>
<li>Needs human input mid-process</li>
</ul>
<h2>Quick Reference</h2>
<h3>Agent Frontmatter Template</h3>
<pre><code class="language-yaml">---
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
---
</code></pre>
<h3>Fix Bash Approval Spam</h3>
<ol>
<li>Remove Bash from tools if not needed</li>
<li>Put critical instructions FIRST (right after frontmatter)</li>
<li>Use allowlists in <code>.claude/settings.json</code></li>
</ol>
<h3>Memory Crash Recovery</h3>
<pre><code class="language-bash">export NODE_OPTIONS=&quot;--max-old-space-size=16384&quot;
source ~/.bashrc &amp;&amp; claude
</code></pre>

Agent Development

AI Skill Market Insights

Be Part of the 0+ Developer Community