Tree of thoughts prompting is a way to explore more than one plan before you commit. Instead of asking the model to “think step by step” down a single path, you ask it to generate several intermediate options, evaluate them, and continue with the better branch. In Tree of Thoughts: Deliberate Problem Solving with Large Language Models, Shunyu Yao and colleagues at Princeton and Google DeepMind described this as a generalization of chain-of-thought prompting. A “thought” is a coherent unit of text that can be a step toward the answer: a partial plan, a candidate outline, or a short writing strategy. The model can look ahead, score options, and backtrack if a branch is weak.
The paper’s most cited result is not a writing test. On Game of 24, a puzzle that uses four numbers and basic arithmetic to reach 24, GPT-4 with chain-of-thought prompting solved 4% of the tasks, while their Tree of Thoughts method reached 74%. Those figures are from the paper’s Game of 24 experiments, not from eBook work. They are a reason to try branching on planning problems, not a promise about your chapter quality.
The authors also tested a creative writing task that is closer to book work. The model had to write a coherent four-paragraph passage that ended on four given sentences. They sampled several plans, voted for the best one, then sampled and voted on passages. A GPT-4 automatic score averaged 7.56 for Tree of Thoughts, against 6.93 for chain-of-thought and 6.19 for a direct “input-output” prompt, on a 1-10 coherency scale. In a blind comparison of 100 passage pairs, human judges preferred Tree of Thoughts over chain-of-thought 41 times, preferred chain-of-thought 21 times, and called 38 pairs similarly coherent. Again, that is the paper’s writing task, not your book.
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 you can compare outline options in a planning stage and keep drafting on the plan you actually accept. The Prompt Engineering Lab in the Writers Workshop is a place to run one linear brief and a branched brief on the same notes, then compare the plans. 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 also not the same move as prompt chaining. Chaining splits a project into separate jobs you can inspect: plan, then draft, then edit. Tree of thoughts explores alternative paths inside a planning job. You can use both. In practice, you branch while you plan, then chain the accepted plan into a draft.
How is tree of thoughts different from chain of thought?
Chain-of-thought prompting produces one sequence of intermediate steps, then an answer. That is efficient when the job has a fairly clear path: check a draft against your notes, or turn an approved outline into a section. The weakness, which Yao et al. point out, is that token-by-token generation commits early. If the first step is a weak chapter angle, the rest of the chain often follows it.
Tree of thoughts keeps several candidates in play. At each step the model proposes more than one thought, evaluates them, and expands the more promising ones. The original research setup is closer to a search algorithm than to a single chat message: it may use breadth-first or depth-first search, a beam of candidates, and a separate evaluation prompt. In an ordinary chat window you can still use the idea in a lighter form, the same pattern the paper used for creative writing: sample a few plans, vote or score them, then write from the winner.
| Approach | What it does | When it helps an eBook |
|---|---|---|
| Direct prompt | One answer, no plan | A small, already-defined page |
| Chain of thought | One linear reasoning path | Checking constraints or structuring one exercise |
| Tree of thoughts | Several paths, scored, with optional backtracking | Choosing a book or chapter structure when the first idea may be wrong |
| Prompt chaining | Separate requests for separate jobs | Moving from an accepted plan to a draft, then to an edit |
Yao et al. recommend Tree of Thoughts for tasks that need deliberate reasoning, where chain-of-thought struggles. They also note the cost: more model calls, more tokens, and a performance-cost tradeoff you can change by shrinking the number of branches. In their creative writing analysis, Tree of Thoughts used about five times as many completion tokens as a simpler pass. For a book, that is an argument for branching at the outline stage, not for every paragraph.
When should you branch an eBook outline, and when should you not?
Use a branched pass when an early structural choice is hard to undo: the order of chapters, the promise of the book, or whether a workbook should teach one skill or three. A single chain can still be the right tool when you are validating a claim against notes or polishing a section whose shape is already settled.
A simple rule: if you would be annoyed to discover, two chapters later, that you picked the wrong frame, branch now. If you already approved the frame, do not spend a tree on it.
Example 1: compare three outlines before you draft
A one-path prompt is:
Here are my notes. Think step by step and write the best outline
for a short eBook that helps new freelance designers name a floor rate.
That may produce a fluent outline. It will not show you the options it discarded, and it may hide a weak opening inside a tidy list of headings.
A tree-style planning pass, in two chat turns, looks like this.
Turn 1: propose and score plans, but do not draft.
I am planning a short eBook for new freelance designers who undercharge.
The book should help them name a floor rate before the next inquiry,
not build a full pricing menu.
Notes:
- The reader currently charges favors or student-level prices.
- I have a $90 floor example for a one-page logo revision.
- I do not have survey data and I do not want invented averages.
- Each chapter should end with one action the reader can complete.
Propose three genuinely different outlines, not three rewordings.
For each outline, include:
- A one-sentence promise
- Five chapter headings
- What this plan does well
- What it risks (scope, repetition, missing the floor-rate job)
Then score each plan 1-5 on:
- Fit to the reader's job
- Specificity (does it stay inside my notes?)
- Completeness of the workbook actions
Do not write chapters. Do not invent a winner if two plans are close;
tell me what I still need to decide.
Turn 2: once you have picked a plan, or asked a follow-up, lock it in.
Use Plan B, with this change: move the "say the floor on the call"
chapter earlier, and drop any heading that would require industry
averages I do not have.
Return a final five-chapter outline, a one-line job for each chapter,
and the workbook action at the end of each chapter. Then stop.
The important part is the pause between turns. The paper’s method evaluates thoughts before it continues. If you let the model pick a winner and start drafting in the same breath, you are back to a single chain with extra decoration.
Example 2: branch a single overloaded chapter
Sometimes the book structure is fine and one chapter is the problem. Suppose your “pricing” chapter is trying to cover a floor rate, a full package menu and a discount policy at once.
This chapter is doing three jobs at once, and I need to pick one.
Reader: new freelance designer. Book job: name a floor rate.
The three candidate jobs for this chapter are:
A. Help the reader name a floor rate and say it on the next call.
B. Help the reader build a three-tier package menu.
C. Help the reader write a discount policy.
For each option, give:
- A five-heading mini-outline
- Which of my notes it uses
- What would have to move to another chapter or out of the book
Then recommend one option for this book, or say if the book should
be split. Do not blend A, B and C into a single "comprehensive"
chapter.
A chain-of-thought prompt will often “resolve” this by combining all three. The branched brief makes the combination visible as a choice, which is the point.
Example 3: a cheap vote when you already have two outlines
You do not always need five branches. If you already wrote two outlines, or the model produced two in earlier chats, you can run a scoring pass on your own.
Here are two outlines for the same short eBook. I will pick one.
Score each outline 1-5 on fit to the reader's job, specificity,
and risk of inventing material I do not have. Quote the heading
that most helped or hurt each score. Then tell me the main
trade-off I am choosing between. Do not rewrite the outlines.
This is closer to the paper’s “vote” step than to a full search tree. It is also easier to inspect, because you can disagree with a score.
How does this fit a planning stage in a book workspace?
Spec Builder is the natural home for a branched outline. Generate two or three plans, compare them against the reader and your notes, and only then let Drafter work from the accepted structure. Polisher should not be inventing a new book shape while it edits. If you use the Prompt Engineering Lab, treat “one outline path versus three scored paths” as a single-variable test: keep the notes fixed, change only whether the brief asks for branches, and compare the plans on a normal brief and on a messy one that tries to cover too much.
Keep the search shallow. Yao et al. used a depth of two for their creative writing task: pick a plan, then pick a passage. For an eBook, “propose a few outlines, score them, finalize one” is usually enough. A deeper tree that re-plans every section will spend tokens and still need your judgment.
FAQ
What is tree of thoughts prompting?
It is a prompting and inference method from Yao et al. (2023) that explores several intermediate thoughts, evaluates them, and continues down the more promising branches, with optional backtracking. In an eBook context, a thought is usually a plan, an outline or a chapter angle, not a finished chapter. The research setup can involve many model calls. In a chat window, the useful version is: generate a few distinct options, score them, then proceed with the one you accept.
How is tree of thoughts different from chain of thought?
Chain of thought follows one reasoning path to an answer. Tree of thoughts keeps multiple paths in play and compares them. Chain of thought is usually enough when the job is already framed. Tree of thoughts is for decisions where the first frame might be wrong, such as the structure of a book. It is heavier, so use it at the planning stage rather than in every drafting prompt.
How do I use tree of thoughts in ChatGPT?
You will not get the paper’s full search algorithm from one casual message. Use two or three turns: first ask for several genuinely different plans and a score for each, then choose or adjust a plan yourself, then ask for the final outline or draft. Do not let the model declare a winner and write the book in the same reply. If a reasoning model is doing the planning, keep the brief focused on options and criteria, and avoid a long “think step by step” script. Check the current docs for the model you are using.
Before you hit send
- Is this a structural decision, or a job whose shape is already approved?
- Have you asked for genuinely different options, not three paraphrases?
- Did you include a scoring rubric that matches the book’s actual job?
- Will you inspect the comparison and choose, instead of accepting an automatic winner?
- Are you stopping at the plan, then drafting in a later pass?
Branching is a planning tool. Once the plan is good, a straight draft is usually the next step, not a bigger tree.