mcpField
M
multitask

multitask

github.com/owainlewisAgent workflows42379 installsUpdated 2d ago
Repository

About multitask

multitask is a skill for coordinating work across multiple tickets using separate agent workers. It groups overlapping tasks, assigns isolated worktrees, monitors progress, and reports each ticket’s outcome.

Key features

  1. 01Group tickets by overlapping changes
  2. 02Assign workers separate worktrees and branches
  3. 03Delegate implementation through PR and ticket updates
  4. 04Monitor workers and report failures
  5. 05Summarize status, proof, and blockers
  6. 06Limit parallel workers to around five

Use cases

  1. 01Implement several independent tickets concurrently
  2. 02Separate tickets that affect overlapping files
  3. 03Track worker results across branches and PRs

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/owainlewis/blueprint --skill multitask

After installing, type/multitaskin Claude Code to confirm it loaded.

Details

From the project README

Multitask

  1. Read each ticket and the code it likely touches. Tickets that overlap in files, data shapes, or behavior go in one ordered group; the rest get separate workers.
  2. Give each worker its own worktree, branch, and full session. If parallel workers are not available, run groups sequentially and say so.
  3. Start each worker with a complete prompt: ticket, repo path, branch rule, acceptance criteria, checks, review rule, PR rules, report format. The worker owns the full loop from implement through PR, ticket update, and proof.
  4. Watch the workers. If one fails or invalidates an assumption another needs, stop the affected workers and report why. Report failures, never silently retry.
  5. Report every ticket: branch, PR URL, status, blockers, proof, and any worktree path. Clean up worktrees only when the repo convention says to or the user asks.

Rules

  • A worker is a Codex thread or a Claude Code subagent, one per ticket, each in its own worktree.
  • Worker reports are short: branch, PR URL, status, proof. Nothing else returns to the coordinator.
  • Cap parallel workers around five.
  • One ticket, one branch, one PR. Never merge, rebase, or cherry-pick between workers.
  • The coordinator never edits code. Fixes happen inside the worker.
  • A failed worker is reported, not folded into another worker's task.

Comments

Markdown supported