agent-creator

Create a markdown file in .claude/agents with YAML frontmatter containing name and description fields, followed by a system prompt body. Run init_agent.py to scaffold a template, fill in the TODO placeholders, then validate with validate_agent.py.

Always specify the tools field explicitly, since omitting it inherits all tools. Use Read, Grep, Glob for read-only research agents, and add Write, Edit, Bash for implementation agents. Grant only what the agent's domain requires.

Use haiku for exploration and simple search tasks, sonnet for code review and implementation, and opus for architecture decisions and complex reasoning. The inherit option matches the user's session model and is the recommended default.

Auto-triggering depends entirely on the description field, which Claude reads to decide delegation. Write action-oriented triggers like 'Use PROACTIVELY when...' with specific task scenarios instead of vague phrases like 'helps with databases'.

Run validate_agent.py with the path to your agent markdown file or agents directory. It checks frontmatter format, required fields, description quality, tool and model validity, and body structure conventions, reporting errors, warnings, and suggestions.