Step-back prompting is a two-step method: first ask for a higher-level principle or concept, then use that abstraction to do the original job. In Take a Step Back: Evoking Reasoning via Abstraction in Large Language Models, Huaixiu Steven Zheng, Swaroop Mishra and colleagues at Google DeepMind argued that models often stumble on detailed, multi-step questions because an intermediate step goes wrong. Their fix was to “step back” and retrieve a broader concept first, then reason from that concept to the original question.
The paper’s results are from knowledge and reasoning benchmarks, not from book writing. On PaLM-2L, they reported gains of 7 percentage points on MMLU Physics, 11 points on MMLU Chemistry, 27 points on TimeQA, and 7 points on MuSiQue, and they also tested GPT-4 and Llama 2 70B. Treat those figures as evidence that abstraction can help on hard, detail-heavy questions, not as a forecast of your chapter quality.
The running examples in the paper are easy to translate into teaching books. Instead of asking which school a person attended in a specific year, they first ask for the person’s education history. Instead of jumping into a physics word problem, they first retrieve the relevant principle, such as the ideal gas law. The step-back question is supposed to be easier and more general than the original one.
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. A step-back pass belongs in planning: settle the principle, then write the chapter from it. In the Prompt Engineering Lab, you can change only the first instruction, predict how the outline will change, and test the brief on a clean topic and on a pile of over-detailed 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 related to, but not the same as, chain-of-thought prompting. Chain of thought walks through the original problem step by step. Step-back prompting changes the first question to a broader one, then returns to the original job. If your notes are already messy, you may also need a cleaner brief before either method will help.
What does a good step-back question look like?
Zheng et al. define a step-back question as a derived question at a higher level of abstraction. It should retrieve the concept or first principle that the later reasoning will use. It should not be a vague invitation to “think about the big picture,” and it should not be so broad that any principle would count.
For an eBook chapter, a useful step-back question often takes one of these shapes:
- What is the rule or principle the reader must understand before they can do this task?
- What is the category of problem this chapter is really about?
- What does a complete version of this kind of page need, before we fill in this week’s details?
A weak step-back question restates the chapter title, or asks for a summary of all your notes. That is not abstraction. That is more detail.
| Original request | Better step-back question | Why it is broader |
|---|---|---|
| Write the chapter on saying a $90 floor on a sales call | What is the difference between a floor rate and a full pricing menu? | It names the concept the call script depends on |
| Write a workbook page for a first weekly review | What makes a weekly review completable in 15 minutes? | It asks for the design principle, not the page |
| Explain why my 2024 tax example should go in chapter 4 | What job is chapter 4 supposed to do in this book? | It lifts the decision out of one anecdote |
| Draft a quiz about all six tools in my notes | What should this quiz certify that the reader can do? | It defines the skill before the item list |
The second step is then ordinary, constrained writing: use the principle you just agreed on, plus only the notes that support it.
When does this help a how-to chapter, and when does it fail?
Step-back prompting helps when the draft keeps getting lost in particulars: too many tools, too many exceptions, or a story that is more specific than the lesson. It also helps when you are not sure whether a chapter is about a principle or about a single tactic.
It fails when you do not have a reliable principle to retrieve. If the model invents a “universal rule” that is not in your notes, the second step will sound coherent and still be wrong. It also fails when the chapter is already a narrow, approved task, such as writing a 300-word section from a fixed outline. In that case, stepping back adds a delay and a chance to wander.
Always inspect the abstraction. If the principle is off, do not let the model write from it.
Example 1: retrieve the principle, then outline the chapter
A direct prompt piles the details into one request:
Write a chapter that tells new freelance designers how to say a $90
floor on a discovery call, how to handle a discount request, how to
talk about rush fees, and how to follow up if the client goes quiet.
That brief invites a grab-bag chapter. A step-back version splits the work.
Turn 1: abstraction only.
I am writing a short eBook for new freelance designers who undercharge.
The working chapter title is "Say the floor on the call."
Before any outline, answer this step-back question:
What is the principle the reader must understand before they can say
a floor rate out loud?
Use only these notes:
- The book is about naming a floor, not building a full pricing menu.
- The only figure I will publish is a $90 floor for a one-page logo
revision.
- I do not have data on "typical" discounts or rush fees.
Return:
- The principle, in one or two sentences
- What the chapter should not try to teach, if it follows that principle
- Any gap I still need to fill
Do not write the chapter.
Turn 2: reason from the accepted principle.
Use this principle, which I have accepted:
"A floor is the lowest fee you will accept for a defined job. It is
not a catalog of packages, discounts or rush add-ons."
Now write a five-heading outline for the chapter "Say the floor on
the call," plus one workbook line. Stay inside the principle and the
notes. If rush fees or discount policy would need their own chapters,
say so instead of squeezing them in.
The value is the checkpoint. You can reject the principle, tighten it, or notice that the model added a rule you do not believe.
Example 2: step back from an over-detailed notes dump
Teaching books often start as a pile of exceptions. A step-back question can recover the lesson.
Here is a messy set of notes for a chapter on revision limits.
Do not write the chapter yet.
Notes:
- I once did eight rounds on a logo and regretted it.
- I want a default of two revision rounds in the kickoff email.
- One client needed a third round because the printer changed a
color profile; I would make that exception again.
- I also have thoughts about fonts, contracts, and late invoices.
Step-back question: What is the principle for revision limits in
this book?
Then, only after you state the principle, list:
- Which notes belong in this chapter
- Which notes belong somewhere else
- Which notes are anecdotes that illustrate the principle, not
extra rules
If the printer exception is truly an exception, keep it labeled as
one. Do not turn it into a second policy.
If the model answers with a vague slogan such as “set clear boundaries,” ask it to try again until the principle is specific enough to accept or reject.
Example 3: use a step-back question on a quiz or workbook page
Abstraction is also useful when a worksheet tries to test everything.
I need a one-page workbook exercise for the chapter on floor rates.
Step-back question first: What should a reader be able to do, in
concrete terms, after they complete this page?
Then design one exercise that tests only that skill. Include:
- A one-sentence instruction
- The fill-in lines
- One worked example that uses my $90 floor
- One "if you get stuck" prompt
Do not add a second exercise about discounts, contracts or portfolios.
This keeps the page teachable. The principle is the skill. The page is the practice.
How should you keep the two steps honest?
Write them as two passes, even if you paste both into one chat. Read the principle as if it were going in the book. If you would not defend it, do not draft from it. If the principle depends on a source, supply the source; do not let the model “retrieve” a law, study or “first principle” you cannot point to.
In a workspace, Spec Builder can store the accepted principle next to the chapter job. Drafter then writes from that principle and the notes, not from a fresh abstraction of its own. Polisher can check whether the draft still matches the principle, or whether a late anecdote has taken over the lesson. In the Prompt Engineering Lab, change only the opening request: one version asks for the chapter immediately, the other asks the step-back question first. Predict whether the outline will be narrower, then test it on a tidy brief and on a notes dump that includes tempting off-topic detail.
FAQ
What is step-back prompting?
It is a two-step prompting method from Zheng et al. (2023). First the model answers a broader, easier question that retrieves a concept or principle. Then it uses that abstraction to answer the original question. For eBook writers, the first step is often “what principle is this chapter teaching?” and the second step is the outline or draft.
How is step-back prompting different from chain of thought?
Chain of thought works through the original problem in intermediate steps. Step-back prompting changes the first question to a higher-level one, then returns to the original job. You can combine them, but they are not the same request. If a reasoning model is doing the work, keep the first question clear and avoid a long script of how to think; check the current vendor guidance for that model.
How do I use step-back prompting for writing?
Write the step-back question so it names a principle, a category or a skill, not a summary of your notes. Inspect the answer. Then, in a second pass, ask for the outline, chapter or worksheet, and require the model to stay inside the accepted principle and your sources. If the first answer is vague or invented, fix it before anyone drafts.
Before you hit send
- Is the chapter actually stuck on details, or is the job already clear?
- Is your step-back question broader and easier, or just fuzzier?
- Can you defend the retrieved principle from your notes or sources?
- Are you inspecting the abstraction before the draft?
- Have you told the model what the chapter should refuse to cover?
Step back to get a rule you believe. Then write the page from that rule, not from every note you happen to have.