Interactive eBook BuilderInteractive eBook Builder
$59 · unlimited access, unlimited eBooksGet it for $59

· By · 10 min read

Least-to-Most Prompting: Break a Book Into Subproblems

promptsaiwritingebook

Least-to-most prompting has two stages: break a hard problem into simpler subproblems, then solve those subproblems in sequence, using earlier answers as context for later ones. In Least-to-Most Prompting Enables Complex Reasoning in Large Language Models, Denny Zhou and colleagues at Google Research argued that chain-of-thought prompting often struggles when the test problem is harder than the examples in the prompt. Their method aims at that easy-to-hard gap: the model first decomposes the task, then works from the easier parts up to the original question.

The paper’s headline result is from a compositional benchmark, not from writing. Using GPT-3 code-davinci-002, least-to-most prompting solved the SCAN benchmark in any split, including the length split, with at least 99% accuracy from only 14 exemplars, compared with 16% accuracy for chain-of-thought prompting on that setup. The authors note that some specialized neural-symbolic models trained on more than 15,000 SCAN examples were being compared against a short prompt. That is evidence the method can help with problems built from smaller parts. It is not a claim about your chapter quality.

For eBook and workbook projects, Interactive eBook Builder is a $59 one-time Notion AI template with Spec Builder, Drafter, Polisher and SEO BOT already set up, so a hard chapter can be decomposed in planning and only then drafted. The Prompt Engineering Lab can compare one oversized “write this chapter” prompt with a decomposed sequence on the same notes. Notion AI access and usage are separate from the purchase. The examples below are teaching prompts written for this article. They are not paid Builder prompts, and they are not recorded model outputs.

This is easy to confuse with prompt chaining. Chaining is a pipeline of different jobs: plan, then draft, then edit. Least-to-most stays inside one hard job and splits it into easier questions whose answers accumulate. You can use chaining around it: decompose and solve in a planning chain, then draft the chapter in a later chain. It is also a complement to prompts for long-form content and to a fixed prompt structure for turning an outline into a draft. Those articles tell you to work section by section. Least-to-most tells you which questions to settle before a section is ready to write.

How do you tell least-to-most apart from nearby methods?

MethodWhat gets splitWhat moves forward
Chain of thoughtIntermediate steps inside one answerA single path of reasoning
Prompt chainingDifferent jobs (plan, draft, edit)The output of one stage becomes the next stage’s input
Tree of thoughtsAlternative paths at the same decisionA scored choice among branches
Least-to-mostEasier subquestions of the same hard jobEach answer is available when you solve the next, harder part

The giveaway is the order: easier first, and later prompts quote the earlier answers. If you jump straight to “write the chapter,” you are not doing least-to-most, even if you used several messages.

When should you decompose a chapter?

Decompose when the chapter is hard because it depends on smaller questions you have not answered: which figure to publish, what the reader must do, what to leave out, or how the exercise will work. Do not decompose a section that already has an approved job, length and notes. In that case, a single drafting brief is enough, and extra subquestions become delay.

A useful test: if you cannot write the heading in one sentence without a “and also,” you probably have more than one subproblem.

Example 1: decompose a chapter, then solve it in order

A too-hard prompt is:

Write a chapter that teaches new freelance designers how to name a
floor rate, say it on a call, handle discounts, and close with a
workbook page, using my messy notes.

That request is several chapters wearing one coat. A least-to-most sequence looks like this.

Turn 1: decompose only.

I need to plan one chapter for new freelance designers: "Name a floor
rate you can say out loud."

Break this chapter into the smallest list of subproblems I must
solve before anyone can draft it. Order them from easiest to hardest.
For each subproblem, say what a complete answer would look like.
Do not answer them yet, and do not write the chapter.

Suppose the model returns a list such as: define a floor; choose the example figure; write the call script constraints; design the workbook close; then draft. You now solve them in that order.

Turn 2: easiest first.

Subproblem 1: In one or two sentences, define a floor rate for this
book. Use only this note: a floor is the lowest fee the reader will
accept for a defined job, not a full pricing menu. Then stop.

Turn 3: feed that answer forward.

Accepted definition: "A floor is the lowest fee you will accept for
a defined job, not a catalog of packages or discounts."

Subproblem 2: Which figure should the chapter use as an example?
Notes: one note says $75, another says $120, and I have no survey.
If you cannot choose without inventing a compromise, list this as
"Need from me" and stop.

If it flags the conflict, you decide. Only then do you continue to the script, the workbook line and, last, the draft. The last prompt should include the accepted definition, the chosen figure, and the workbook line, so the hard writing step is no longer also a decision step.

Example 2: build a worksheet from easier pieces

A workbook page often fails because the model is asked to invent the skill, the example and the layout at once.

Decompose this page into subproblems, easiest first:
"A one-page exercise that lets the reader name a floor for a
one-page logo revision."

Then solve them in order. After each answer, wait.
The last step should be the finished page, and it must use only
answers I have accepted.

A sensible order is: the skill being practiced; the single worked example; the fill-in lines; then the assembled page. If you skip to the assembled page, the model may add a second exercise “for completeness.”

Example 3: when not to decompose

If the outline, figure and close are already approved, do not add a least-to-most ritual.

Write the "Name the floor before the call" section.

Reader: new freelance designer.
Length: 280 to 320 words.
Use the accepted definition and the $90 example already below.
End with the fill-in line I supplied. Do not add a discount section.

That is a drafting brief. Extra subproblems would only invite the model to reopen decisions you have already made.

What does a short book-level decomposition look like?

You can use the same method at the book level, once, before chapter drafting begins. Keep the list coarse.

  1. Who is the reader, and what will they be able to do at the end?
  2. What is the one job of the book, and what jobs will you refuse?
  3. What examples and figures are you actually willing to publish?
  4. What is the five-chapter sequence that follows from those answers?
  5. Only then, decompose any chapter that is still doing too much.

Questions 1 to 3 are easier than question 4. If you start with “write a 12-chapter outline from these notes,” you have skipped the easy parts, and the outline will invent scope. Store the answers to 1-3 as standing material, then use chapter-level least-to-most only where a chapter still contains an open decision.

How should this live in a writing workspace?

Spec Builder is the place to store the subproblem list and the accepted answers. Drafter should see the accumulated answers, not the original messy request. Polisher can check that the draft still matches those answers and has not reopened a closed decision. In the Prompt Engineering Lab, compare one prompt that says “write the whole chapter from these notes” with a decomposed sequence on the same notes. Use a failure case that contains a conflict, such as two different prices, and see which approach flags it.

Keep the chain short. Three to five subproblems is plenty for one chapter. If the list looks like an entire book, you are decomposing the wrong unit.

FAQ

What is least-to-most prompting?

It is a method from Zhou et al. (2022) that first breaks a hard problem into simpler subproblems, then solves them in sequence, with each earlier answer available for the next step. The aim is to help a model generalize from easier parts to a harder whole. In eBook work, the “hard problem” is often a chapter or a worksheet that still contains unresolved decisions.

How is least-to-most different from prompt chaining or chain of thought?

Chain of thought is one path of intermediate steps inside a single answer. Prompt chaining is a pipeline of different jobs, such as plan, draft and edit. Least-to-most is a particular way of splitting one hard job: easier questions first, then harder ones that reuse those answers. You can put a least-to-most sequence inside a planning chain, then draft in a later chain.

How do I break a writing task into smaller prompts?

List the questions that must be answered before a draft is possible, and order them from easiest to hardest. Solve one at a time. Paste the accepted answers into the next prompt. If a question needs a decision only you can make, stop there. When the remaining job is just writing from settled material, switch to an ordinary drafting brief.

Before you hit send

The last prompt should be easy because the hard questions are already settled. If it is still hard, you skipped a subproblem.

Ship niche eBooks with agents already built.

Get the Interactive eBook Builder for $59