Agent Development

This skill provides a comprehensive framework for developing autonomous agents within Claude Code plugins. It is designed for developers who need to implement multi-step, independent subprocesses that operate alongside standard user-initiated commands. The skill defines the architectural requirements for agents, including the usage of Markdown-based definitions with YAML frontmatter, strict naming conventions, and specific model and tool access configurations. By utilizing this framework, developers can create robust agents capable of handling complex domains, from security analysis to automated code generation.

• Define autonomous behaviors using second-person system prompts that specify roles, responsibilities, analysis processes, and output formats.

• Implement precise triggering conditions using the description field, ensuring agents activate only when contextually appropriate based on provided examples.

• Utilize YAML frontmatter to manage agent metadata, including required model selection (inherit, sonnet, opus, or haiku), visual identifiers (colors), and tool restriction sets for security and least-privilege access.

• Leverage expert-designed AI-assisted generation prompts to bootstrap agent configurations, ensuring adherence to the mandatory agent structure.

• Always follow the required naming pattern: lowercase, numbers, and hyphens only, with a length between 3 and 50 characters.

• Include multiple concrete example blocks in the description to guide the LLM on exactly when to trigger the agent, including context, user requests, assistant responses, and reasoning commentary.

• Adhere to tool-limitation best practices; specify explicit tool arrays (e.g., Read, Write, Grep) rather than providing blanket access to ensure system integrity.

• Maintain system prompt brevity by keeping text under 10,000 characters and focusing on clear, step-by-step methodologies to handle tasks and edge cases effectively.

• Use this skill when building plugins for code review, automated testing, API documentation, or any scenario requiring persistent, specialized autonomous logic.