parallel-execution_skill

Overview

This skill provides patterns and concrete steps for launching and coordinating parallel subagent execution using the Task tool with run_in_background. It focuses on spawning dynamic subagents in a single assistant message, collecting TaskOutput, and synthesizing results into a unified outcome. Use it to speed up independent, non-conflicting work by running many tasks concurrently.

How this skill works

The skill inspects a planned workload, identifies independent tasks or directories that can be parallelized, and generates per-subagent prompts describing role, files, and expected output. It instructs the orchestrator to launch all Task calls in one assistant message with run_in_background: true, then retrieve each TaskOutput once tasks finish, and synthesize the outputs into a final deliverable. It includes patterns for task-based, directory-based, and perspective-based parallelization, plus guidance for TodoWrite status handling.

When to use it

• Multiple independent analyses (security, performance, tests)
• Multi-file processing where files or components are independent
• Implementing several independent feature components concurrently
• Exploratory reviews from different perspectives (architecture, UX, QA)
• Verification or linting checks that do not share state or files

Best practices

• Place every Task call in the same assistant message to ensure true parallelism
• Verify tasks are independent: no shared file modifications or required outputs from other tasks
• Write explicit, focused subagent prompts with role, scope, files, and expected output format
• Use unique task IDs and TaskOutput retrieval to avoid race conditions
• Reserve a final synthesis step that merges findings, resolves conflicts, and produces actionable results

Example use cases

• Implementing an auth feature by spawning subagents for login, registration, middleware, password reset, and tests
• Analyzing separate directories (src/auth, src/api, src/db) in parallel for issues and suggestions
• Performing simultaneous reviews from Security, Performance, Testing, and Architecture perspectives
• Running multiple verification checks (lint, typecheck, unit tests) across independent modules
• Spawning one subagent per planned implementation task in a project plan