Break It Down: Least-to-Most, Plan-and-Solve and Tree of Thoughts

Study guide

Break Down Big Jobs for Better AI Results

This course teaches you how to stop asking AI for everything at once. When you paste a giant prompt, the AI spreads its attention thin and gives you a shallow answer. Instead, you will learn to break the job into smaller, focused steps. You will meet five practical frameworks: least-to-most, plan-and-solve, skeleton of thought, tree of thoughts, and decompose and delegate. Each one gives you a clear way to control the pace and quality of the AI's work.

You will see how each framework works in a real job. An accountant reconciles a messy account by starting with the simplest question. A manager approves a plan before any work begins. A writer agrees an outline before expanding ten pages. A strategist explores three branches and picks the best. By the end, you will have a repeatable routine: ask for the step list, approve it, run each step as its own prompt, and save the list as a template. This course is for busy professionals who want deeper, more useful answers from ChatGPT, Claude or Gemini without learning technical jargon.

What is a prompt framework?

A prompt is the request you type into an AI tool such as ChatGPT, Claude or Gemini. A prompt framework is a reusable structure for that request: a short checklist of the parts a good prompt contains, usually named with an acronym so it is easy to remember. Why it helps: the AI fills in whatever you leave out with generic guesses. A framework makes sure you include the parts that change the answer, such as who it is for, what you want to achieve and the format you need.

Frameworks make prompts repeatable: you fill the same parts each time, save the prompt as a template and share it with colleagues. They are not magic words: the value comes from the detail you put in each part. They help most when a request is complex or open to interpretation; a quick, simple question can stay short. There are well over sixty named frameworks, and most of them remix the same ingredients: role, task, context, audience, format, constraints and examples.

In this course you learn Decomposition, which is a prompting technique from AI research: instead of slots to fill, it changes how the AI works through a task, and you can add it on top of any fill-in framework. It is best for: big, complex jobs: plans, long documents and decisions with several possible routes.

Why big prompts fail

A single prompt for a large job forces the AI to spread its attention across everything. Every part gets a surface-level answer. Decomposition fixes this by splitting the job into smaller, focused steps. Decomposition is not a workaround. The Prompt Report, 2024, the largest survey of prompting techniques, lists it as one of six main families of prompting methods.

Each smaller prompt should have a clear role, a clear job and a clear scope for that one step. This stops the AI from guessing what matters and keeps the answer focused. The output of one step often becomes the input for the next. For example, a list of project phases becomes the basis for listing tasks inside each phase.

A common mistake is asking for the whole job at once and then blaming the AI for a generic answer. The prompt asked for too much, so the answer had to stay shallow. Decomposition is the foundation for the techniques in this course, including least-to-most, plan-and-solve, skeleton of thought and tree of thoughts.

Least-to-most: start with the easiest step

Least-to-most prompting was described by Zhou and colleagues in 2022. It lists the sub-problems first and solves the easiest one first, rather than asking for the whole answer in one go. The core prompt is short: break this into sub-problems, ordered from simplest to hardest, and solve the first only. The model does the ordering, and you control the pace.

Each answer feeds the next. The easiest answer becomes a known fact, so the next, harder sub-problem stands on solid ground instead of guesswork. A common mistake is to list the sub-problems and then ask for all of them at once. That defeats the purpose, because no answer can feed the next one. Another mistake is to start with the hardest sub-problem. The hardest part usually depends on the earlier answers, so it needs the simpler steps first.

In the worked example, an accountant reconciling a messy account gets a small first answer, the transaction count and date range, which she can check in seconds before moving on. That small check builds confidence and often reveals a data problem early, before any calculations are done.

Plan-and-solve: approve the approach first

Plan-and-solve comes from Wang and colleagues, 2023. It is a prompting pattern where the AI produces a plan before it produces any output, and you review that plan first. The practical wording is short: show me the plan first, then stop. After I approve, do step 1. The stop is what turns a plan into a real checkpoint rather than a formality.

The value is early correction. A wrong approach caught at the plan stage costs one sentence to fix. The same wrong approach caught after a full draft costs hours of rework. A good plan-and-solve prompt names the goal, asks for the plan, says what the plan should cover, and states the approval step. Leave out the approval line and the AI will often run ahead.

Plan-and-solve differs from least-to-most. Least-to-most breaks a task into smaller solved pieces. Plan-and-solve keeps the task whole but pauses for your approval before execution. Use it when the task is big enough that a wrong direction would be expensive, such as a campaign, a report, a proposal, or a piece of research.

Skeleton of thought: outline before you write

Skeleton of thought splits one big writing job into two: agree the outline, then expand each heading into full text. The split is what keeps a long document consistent. The method is ideal for long documents such as reports, handbooks and proposals, where a wrong heading early on corrupts every section that follows.

Always edit the skeleton before expansion. Changing an outline is cheap, changing ten pages is not, so the review happens while the fix still takes seconds. Add a stop line to the prompt, for example wait for my approval, so the AI does not write the whole document before you have seen the shape.

Expand one section at a time. Each expansion inherits the approved skeleton, which keeps tone, order and coverage stable across the whole document. The skeleton doubles as a checklist. When the chapter is done, you can confirm every heading has text and nothing has been quietly dropped.

Tree of thoughts: explore options before choosing

Tree of thoughts comes from Yao and colleagues in 2023. It treats reasoning as a search: the AI proposes several branches, evaluates each, and backtracks when a branch fails. The everyday version is three experts who propose approaches, critique each other, and agree on the best. This gives you a small debate instead of a single opinion.

Use it for strategy and decisions with real trade-offs, such as cost against speed or quality against risk. If there is no downside to any option, a simple prompt is enough. Your prompt needs three parts: ask for several distinct approaches, set the criteria for judging them, and ask the AI to critique each branch and recommend one.

The critique is the point. Without it you get three lists and no decision. Ask which branch fails, why it fails, and what the winner does better. Keep the number of branches small, around three. Too many branches makes each one shallow, and the final recommendation becomes vague.

Decompose and delegate in your job

The core routine is three moves: ask for the step list, approve or edit it, then run each step as its own prompt. This keeps the plan visible before any work is done, so mistakes are caught early and cheaply. A single step can be delegated to a specialised tool or skill when one exists. Writing steps stay in the chat, number steps go to a spreadsheet, research steps go to a search tool. The AI holds the plan, the tool does the work.

The step list is a reusable template. Save it and paste it next time the same job comes round, changing only the details. Each run makes the list better and the next run faster. Breaking work down also makes checking easier. A short answer for one step is quick to read and quick to fix. If a step is wrong, you redo that step, not the whole task.

This course chains five frameworks. Least-to-most builds from easy to hard, plan-and-solve plans before solving, skeleton of thought drafts the frame, tree of thoughts explores branches, and decompose and delegate runs the whole thing as a daily habit.

Frequently asked questions