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

CODEBLOCK0

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

1. 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.

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:

CODEBLOCK1

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 or ~/.zshrc:
CODEBLOCK2

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

1. 1. source ~/.bashrc or restart terminal
2. INLINECODE6
3. 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 |

CODEBLOCK3

Declarative Over Imperative

Describe what to accomplish, not how to use tools.

Wrong (Imperative)

CODEBLOCK4bash
grep -r "PLACEHOLDER:" build/*.html
CODEBLOCK5

Right (Declarative)

CODEBLOCK6

What to Include

Include	Skip
Task goal and context	Explicit bash/tool commands
Input file paths

"Use X tool to..." | | Output file paths and format | Step-by-step tool invocations | | Success/failure criteria | Shell pipeline syntax | | Blocking checks (prerequisites) | Micromanaged workflows | | Quality checklists | |

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

Discovery	Where to Capture
Bug fix pattern	Agent's "Corrections" or "Common Issues" section
Quality requirement

Agent's "Quality Checklist" section | | File path convention | Agent's "Output" section | | Tool usage pattern | Agent's "Process" section | | Blocking prerequisite | Agent's "Blocking Check" section |

Test: Would a Fresh Agent Succeed?

Before completing any agent improvement:

1. 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

Anti-Pattern	Why It Fails
"As we discussed earlier..."	No prior context exists
Relying on files read during dev

Agent may not read same files | | Assuming knowledge from errors | Agent won't see your debugging | | "Just like the home page" | Agent hasn't built home page |

Agent Prompt Structure

Effective agent prompts include:

CODEBLOCK7

Pipeline Agents

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

Must Update	What
New agent	"Workflow Position" diagram + "Next" field
Predecessor agent

Its "Next" field to point to new agent |

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

Verification:
CODEBLOCK8

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

CODEBLOCK9

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

CODEBLOCK10

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

CODEBLOCK11

Fix Bash Approval Spam

1. 1. Remove Bash from tools if not needed
2. Put critical instructions FIRST (right after frontmatter)
3. Use allowlists in INLINECODE10

Memory Crash Recovery

CODEBLOCK12