Plan-and-Execute Decomposer

What this prompt does

Separates a complex task into two distinct LLM calls: a planner that decomposes the task into atomic, sequenced steps, and an executor that carries out each step one at a time with full context of the plan and prior results. Prevents the model from skipping steps or losing track of progress on long tasks.

When to use it

• Tasks requiring 5+ sequential actions where order matters
• Agentic workflows with tool calls between steps
• Any task where you need to inspect or modify the plan before execution begins

When NOT to use it

• Simple tasks solvable in one or two steps — planning overhead is not worth it
• When plan revision mid-execution is frequent — use a replanning loop instead

The Prompt

Phase 1: Planner

Phase 2: Executor (run once per step)

You are an executor. Complete the current step of a multi-step plan.

Original task: {{TASK_DESCRIPTION}}

Full plan:
{{PLAN_FROM_PHASE_1}}

Completed steps and results:
{{PREVIOUS_RESULTS}}

Current step to execute: {{CURRENT_STEP}}

Instructions:
1. Execute only the current step listed above.
2. Use the completed steps and results for context if needed.
3. Do not skip ahead or execute future steps.
4. Return the result of this step clearly labeled.

Result of Step {{STEP_NUMBER}}:

Why it's structured this way

"Do not execute. Only plan." — Without this constraint, the planner model begins executing steps inline. Strict role separation keeps planning and execution as distinct calls with distinct prompts.

Atomic step definition — "Self-contained and independently executable" prevents step descriptions that span multiple actions, which confuses the executor about where one step ends and another begins.

Previous results injected into executor — The executor needs prior results to handle data dependencies (e.g., Step 3 uses the URL fetched in Step 2). Omitting this causes the executor to hallucinate prior results.

"Do not skip ahead" — Without this, models sometimes attempt to complete multiple steps in one call, producing untracked intermediate state.

Variations