writing-agents

The writing-agents skill provides a structured workflow for implementing, maintaining, and refining autonomous subagents within a project. Designed for engineering workflows, it enforces a Test-Driven Development (TDD) approach where agents are treated as code that must be validated against representative tasks. By defining clear personas, behavioral guidelines, and tool-use boundaries, this skill helps developers create specialized agents capable of executing multi-turn, autonomous tasks that leverage the project's unique context, codebase, and domain-specific best practices.

• Define project-level agents using standardized frontmatter YAML in the .claude/agents/ directory.

• Implement a RED-GREEN-REFACTOR cycle for agent persona development and task execution.

• Research and integrate domain-specific anti-patterns, security considerations, and best practices into agent configurations.

• Establish coordination protocols for multi-agent workflows, including delegation, scope boundary management, and deferral relationships.

• Gather codebase-specific context by leveraging existing tools like glob, grep, and documentation analysis to ground agent behavior.

• Enable specialized automation for complex domains such as build system engineering, CI/CD pipeline management, Debian/RPM packaging, and testing infrastructure.

• Always prompt for a session restart after modifying agent definitions, as agents are initialized during startup.

• Distinguish between agents for autonomous execution and skills for on-demand guidance.

• Use the Task tool with the subagent_type parameter to invoke specific agents.

• Combine research findings from web searches with project-specific documentation (CLAUDE.md, README.md) to build robust, project-aware personas.

• Document explicit scope boundaries to prevent role creep and ensure effective delegation between subagents.