What problem does it solve? Writing a well-configured Claude Code agent requires choosing the right role name, model, effort, maxTurns, and tool permissions, plus a three-part description that avoids triggering collisions with neighboring agents. This Skill guides that process end to end so the resulting agent definition holds up in real project operation. ## Core Features & Use Cases - Structured interview and project vocabulary discovery: Scans README, CONTRIBUTING, CODEOWNERS, and existing agents to name the role after one that would genuinely exist on the project, then interviews the user about mandate, triggering, tools, and memory. - Matrix-based configuration: Assigns model, effort, and maxTurns from a binding decision matrix that accounts for role difficulty and invocation frequency, with mandatory frontmatter comments justifying every omitted field. - Independent review loop: Submits the finished agent to a clean-context review subagent, autonomously processes blocking findings across rounds, and escalates to the user only on mandate disputes or repeated failures. - Use Case: A user says "I need someone to own release decisions on this project." The Skill discovers the project's role vocabulary, interviews about the release manager's mandate, writes .claude/agents/release-manager.md per the template, and iterates with a reviewer until approved. ## Quick Start Ask the assistant to create a new agent for a role your project needs, for example by saying "create an agent that owns release decisions for this project".
