team_skill

Overview

This skill spawns and coordinates N specialist agents to solve a shared task using Claude Code native team primitives. It decomposes a high-level task into independent subtasks, creates a team and task graph, spawns background worker agents, and orchestrates execution, verification, and iterative fixes until completion or a terminal failure. Optional modifiers let you pick worker type for execution or wrap the run in a persistence/retry loop.

How this skill works

On invocation the lead agent parses parameters (agent count, worker type, task), analyzes the codebase to decompose the request into file- or module-scoped subtasks, and creates a team with a runnable task graph. It spawns N teammates in parallel, assigns tasks, monitors progress via inbound messages and task polling, routes stage-appropriate specialists for planning, PRD, execution, verification and fixes, and persists stage state to support resume, cancel, and retries.

When to use it

• Large code changes that can be split into independent file/module-scoped subtasks
• Automated multi-stage workflows requiring plan → execute → verify → fix loops
• Parallelizing repetitive or multi-file fixes like type errors or build failures
• Scenarios needing specialized reviewers during verification (security, performance)
• Runs that should be resumable or cancelled cleanly across sessions

Best practices

• Keep subtasks file- or module-scoped to avoid merge conflicts and race conditions
• Let the lead pick agents for non-exec stages; only override the team-exec worker when necessary
• Pre-assign owners to tasks to avoid race conditions instead of atomic claiming
• Limit fix-loop iterations with a sensible max to avoid infinite retry cycles
• Monitor stuck workers and reassign after defined age thresholds rather than waiting indefinitely

Example use cases

• Fix all TypeScript errors across a repository by decomposing files and routing build-fixers where needed
• Parallelize UI work: spawn designers for layout subtasks and run a verifier for acceptance criteria
• Run a multi-stage feature implementation that needs planning, product scoping, execution, and security review
• Automate triage and repair of CI build failures by routing build-fixer and debugger agents
• Resume a previously interrupted team run from the last non-terminal stage