What problem does it solve? When an AI agent receives a new task, it often processes everything serially in the main context, wasting time on independent subtasks and polluting the context with bulky search results. This Skill defines decision criteria for when to parallelize work across subagents, when to keep work inline, and how to avoid common pitfalls like biased self-evaluation and silently dropped dispatches. ## Core Features & Use Cases - Parallel Dispatch Criteria: Identifies independent subtasks (multi-scenario evaluation, multi-model comparison, independent searches) that should be dispatched to subagents in a single message. - Context Isolation Rules: Routes large-scale exploration (3+ grep/search queries) to general-purpose or Explore subagents so the main context only receives summaries, and mandates bias-free evaluation of your own outputs via fresh subagents. - Long-Running Batch Handling: Covers background execution with run_in_background, Monitor sandbox limitations (no gh access), and retry strategies when parallel dispatches are rejected by permission classifiers. - Use Case: You receive a request to review 20 repositories. Instead of scanning them one by one, you dispatch parallel subagents per repo, keep only summaries in the main context, and re-dispatch any rejected subtasks with rephrased prompts. ## Quick Start Read this Skill at the start of any new task and ask it to identify which subtasks can be parallelized or delegated to subagents before beginning work.
