Claude Agents Creator

Create a focused worker with a concrete trigger, capability boundary, and usable result. Default to a project-scoped Claude Code subagent for repository work; use Managed Agents when an application needs the hosted API harness.

For repository instructions use agent-md-creator; for reusable Agent Skills use skill-creator; for OpenCode workers use opencode-agent-creator.

Build the definition

1. Establish the host, installed version, scope, task, and completion signal. Use an existing built-in worker when it already fits. Derive routine choices from the request; ask only for missing information that changes the result.
2. Give the worker one responsibility. Put the concrete delegation trigger in a short description; put workflow and output requirements in the body.
3. Choose project .claude/agents/, personal ~/.claude/agents/, or plugin agents/. Keep identities unique; check collision precedence and plugin restrictions in references/frontmatter-reference.md.
4. Set the tools the work requires. For a reviewer that must not write, start with Read, Grep, Glob; shell and MCP access can permit writes, and persistent memory can add write tools. A prose boundary alone does not restrict capabilities.
5. Choose an available model alias or ID. Set model: inherit when matching the parent is intended; an omitted field can resolve through other model settings. Add skills, memory, MCP, hooks, or isolation only when the task needs them.
6. Write the body: role, required inputs, essential steps, completion criteria, and returned evidence. Keep stable conventions in the body or preloaded skills; pass the current task's files, constraints, and observations in the delegation message.