Claude Code Master

Use Claude Code CLI (claude) as a sub-agent for coding tasks — code generation, refactoring, codebase analysis, git workflows, and file editing. Always use print mode (-p) for non-interactive execution.

When this applies

• Agent needs to write, edit, or refactor code
• Agent needs to analyze or answer questions about a codebase
• Agent needs to perform git operations (commits, PRs, reviews)
• Agent needs a second AI opinion on code quality or architecture

Self-Help

Run claude --help for the full command list. Run claude -p --help for print mode flags. Use --verbose to debug issues.

Process

1. Set ANTHROPIC_API_KEY in env before first use
2. Use claude -p "prompt" for single-shot tasks
3. Pipe file content when context is needed: cat file.py | claude -p "review this"
4. Use --output-format json when parsing results programmatically
5. Use --max-turns and --max-budget-usd to bound cost and runtime
6. Chain with other tools: claude -p "write tests for auth.py" > tests/test_auth.py

Essential Commands

Single-shot (print mode)

claude -p "explain what this project does"
claude -p "add error handling to src/api.py"
claude -p --model sonnet "quick question about this code"
claude -p --model opus "complex refactoring task"

With input piping

cat src/main.py | claude -p "find bugs in this code"
git diff HEAD~3 | claude -p "summarize these changes"
cat error.log | claude -p "diagnose this error"

Structured output

claude -p --output-format json "list all API endpoints in this project"
claude -p --json-schema '{"type":"object","properties":{"files":{"type":"array","items":{"type":"string"}}}}' "which files need updating"

Cost control

claude -p --max-turns 3 "quick fix for the typo in README"
claude -p --max-budget-usd 0.50 "refactor the auth module"

Custom system prompt

claude -p --system-prompt "You are a security auditor" "review this codebase for vulnerabilities"
claude -p --append-system-prompt "Always use TypeScript strict mode" "add types to utils.js"

Scoped tools

claude -p --tools "Read" "explain the architecture"
claude -p --tools "Bash,Read,Edit" "fix the failing test"

Decision Rules

• Quick question or small fix → claude -p "prompt" (single shot)
• Need specific model → add --model sonnet or --model opus
• Parsing output in script → add --output-format json
• Expensive task → set --max-budget-usd and --max-turns
• Security-sensitive code → add --system-prompt "You are a security auditor"
• Only reading, no edits → use --tools "Read" to restrict capabilities
• Multiple related tasks → pipe output between calls, don't try to do everything in one prompt

Do

• Always use -p (print mode) — interactive mode blocks the agent
• Set --max-budget-usd for open-ended tasks to prevent runaway costs
• Pipe relevant context in rather than relying on codebase discovery for speed
• Use --model sonnet for simple tasks to save cost, opus for complex reasoning
• Use --output-format json when downstream processing is needed
• Restrict tools with --tools when the task only needs reading or specific capabilities

Don't

• Never run claude without -p flag (blocks on interactive input)
• Never skip --max-budget-usd on open-ended tasks (unbounded API spend)
• Never pipe secrets or credentials into claude (they appear in logs)
• Never use --dangerously-skip-permissions (bypasses safety checks)
• Never assume claude modified files — verify with cat or git diff after