Multi-Agent Orchestrator Prompts (Cursor, Claude Code, Codex)

When One Chat Is Not Enough

Large tasks—refactoring a module, shipping a feature across frontend and backend, or running research plus implementation—break when you dump everything into a single prompt. An orchestrator agent plans the work, assigns scoped sub-tasks to workers, and integrates outputs. You stay the product owner; the AI runs the project management loop.

The Orchestrator Persona

Start by forbidding the model from doing all implementation itself unless the task is trivial.

"You are the lead orchestrator. Do not implement everything yourself. Your job: (1) restate the goal, (2) decompose into sub-tasks with dependencies, (3) assign each sub-task to a worker with inputs, constraints, and definition of done, (4) define how workers report back. Output a wave plan (what runs in parallel vs serial)."

Worker Assignment Block

For each worker, specify scope so agents do not collide on the same files.

"Worker A — Implementation: edit only src/api/auth.ts and tests/auth.test.ts. Deliver diff + test command. Worker B — Docs: update README auth section only. Worker C — Review: read Worker A diff; list security issues; do not edit files."

Handoff Format

Require a consistent report so you can merge without re-reading entire threads.

"Each worker returns: Status (done/blocked), files touched, assumptions, verification steps run, open questions. Orchestrator produces final summary + merged checklist for the human."

Conflict Rules

Add coordination constraints used in real agent systems: check for overlapping paths before parallel work, serialize conflicting tasks, retry once with narrower scope, escalate blockers with a short summary.