parallel-task — Claude Code skill: task orchestration and setup

What parallel-task does

The skill parses a Markdown plan into tasks, including their IDs, requirements, validation steps, and dependencies. You can request selected task IDs; their prerequisites are included too. Without a selection, it works through the whole plan. Tasks whose dependencies are complete are launched together, in waves.

For suitable tasks, the workflow starts with a subagent writing tests and showing that they fail, then moves to implementation and passing checks. If a task is not a good fit for TDD, the agent records why and defines another concrete verification method. Once checks pass, the agent commits locally and updates the plan; the orchestrator reviews the result and can request another attempt. A task needs verification evidence, a commit, and a plan update before it counts as done.

For Claude Code, installation places the skill at ~/.claude/skills/parallel-task/SKILL.md. Start it explicitly with /parallel-task, followed by a plan path and optional task IDs. The workflow permits local commits, not pushes.