Cheatsheet: Planning: breaking a goal into steps

The one idea

Planning is the turn from reacting one move at a time to deciding the shape of the whole job first. The agent breaks a goal into an ordered set of sub-tasks, then executes them.

Why reaction breaks down at scale

A reactive agent sees only the current state and the next move. On a long task that causes three failures:

• Wrong order (it cannot see dependencies)
• Repeated or skipped work (no map of what is done)
• Drift off the goal (nothing holds the overall shape)

A plan is the map that fixes all three.

Planning = decomposition

Break one large goal into an ordered list of smaller sub-tasks. The order captures dependencies (you cannot build the itinerary before you know the dates). Produce the breakdown first, then act.

How an LLM plans

It is a reasoning move, the same skill as “think step by step.” Ask the model for the steps first, not the final answer. Write the plan in a structured, executable form (a numbered list / machine-readable steps), not loose prose, so each step can be followed, checked off, and executed.

Worked example

GOAL: 3-day Tokyo trip under $2000, museums AM, food PM.

PLAN (before acting):
  1. check flight prices for candidate dates
  2. pick dates within budget
  3. book lodging for those dates
  4. draft itinerary (museums AM, food PM)
  5. book food tours that fit

EXECUTE: run each step in order, one loop iteration + tool call each.

Plan-then-execute vs replanning

	Plan-then-execute	Replanning
Plan	Built fully upfront, then run	Revised as results come in
Fits	Predictable tasks	Tasks where a step’s result changes the rest

step 1: check flights, first week of June -> nothing under budget
REPLAN: budget beats date preference -> check second week -> found -> continue

Most capable agents sit between: enough planning to give shape, enough replanning to survive reality.

When planning is worth it

Multi-step interdependent goals: yes. Single-step tasks (“weather tomorrow?”): no, the plan is pure overhead. Match planning to the task, like every other capability in this track.

Pitfalls to dodge

• Planning tasks that do not need it (overhead).
• Treating the plan as unchangeable (it is a hypothesis; replan on contradiction).
• Planning in vague prose (steps must be executable).
• Confusing a finished plan with finished work (the plan is the map, not the journey).
• Over-planning a moving target (plan short horizons, replan often).

Words to use precisely

• Planning: deciding an ordered sequence of sub-tasks before (or alongside) acting.
• Decomposition: breaking one goal into smaller, ordered, achievable sub-tasks.
• Plan-then-execute: build the full plan upfront, then run it.
• Replanning: revising the remaining plan based on a step’s actual result.