Zero-shot prompting gives the model instructions only; few-shot prompting adds a handful of input/output examples so the model can copy the pattern. Use zero-shot when the task is clearly stated; switch to few-shot when format, voice or edge cases keep drifting. In Language Models are Few-Shot Learners, Brown and colleagues at OpenAI defined the distinction as part of in-context learning: the model performs a new task from a text instruction and, when provided, demonstrations, without updating its weights. Zero-shot uses only the instruction. One-shot adds a single demonstration. Few-shot adds several, often as many as will fit in the model’s context window.
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 keep a standing brief and a few approved examples next to the chapter you are writing. Notion AI access and usage are separate from the purchase. The examples below are teaching prompts written for this article, not paid Builder prompts and not recorded model outputs. If you already know you need examples to keep chapters consistent, our guide to few-shot prompting for consistent eBook chapters focuses on that use. This article compares the two setups and helps you decide when the extra examples are worth the space.
What is the difference, in practice?
In a writing workflow, the difference is easy to feel. A zero-shot prompt says what to do. A few-shot prompt also shows what “done” looks like.
Google’s prompt design strategies use the same definitions: prompts with no examples are zero-shot, and prompts with a few examples are few-shot. Google notes that few-shot prompts are often used to regulate formatting, phrasing, scope or general patterning, and recommends specific, varied examples. OpenAI’s prompt engineering guide describes few-shot learning as steering a model toward a new task by including a handful of input/output examples rather than fine-tuning it, and it advises showing a diverse range of possible inputs with the desired outputs.
OpenAI’s help article on prompt engineering suggests starting with zero-shot, moving to few-shot if needed, and considering fine-tuning only after those options have been tried. OpenAI’s reasoning-model guidance makes a similar point for those models: try a prompt without examples first; add a few closely aligned examples only if the output requirements are more complex. Google’s current page, by contrast, recommends always including few-shot examples, and says prompts without them are likely to be less effective. Those two pieces of official advice are not identical. The practical way to handle the disagreement is to test both versions on the same page of your book, using the same notes, and keep the one that meets your requirements.
| Setup | What you provide | Best when | Watch for |
|---|---|---|---|
| Zero-shot | Instructions, notes and a format | The job is clear and you do not yet have a trusted sample | The model invents a default structure or voice |
| One-shot | The same brief plus one demonstration | You have a single approved page and want the next one to match it | The model copies incidental details from that one sample |
| Few-shot | The brief plus a few demonstrations | Voice, structure or edge cases keep drifting | Too many similar examples can over-narrow the result |
Brown et al. treated one-shot as its own setting because some tasks are hard to communicate with instructions alone, and a single demonstration often matches how people are briefed. For an eBook, one approved chapter opening is often enough to start; two or three become useful when chapters have different jobs, such as a teaching chapter, a case chapter and a workbook chapter.
When is a clear zero-shot prompt enough?
Zero-shot is a good first tool when you can state the reader, the job, the notes and the format without pointing at a sample. Anthropic’s prompting best practices start with being clear and direct: say what the output should contain, and use steps when the order matters. If that brief already produces a usable outline, checklist or short explanation, you may not need examples yet.
A vague zero-shot prompt looks like this:
Write the introduction to my chapter on client boundaries.
A clear zero-shot prompt looks like this:
Write a 200 to 250-word opening for a chapter on client boundaries, for
new freelance designers who say yes to extra revisions because they are
afraid of seeming difficult.
Use only these notes:
- The chapter's job is to help the reader state a revision limit before
work starts, not to write a legal contract.
- The reader already uses email, not a formal proposal tool.
- Do not invent laws, rates or "industry standards."
Open with the reader's situation, then state the chapter's job in one
sentence. End with a one-line preview of the three-step method that
follows. Use plain words. Do not tell a celebrity anecdote.
If the result still misses the shape—maybe it writes a manifesto, or it buries the job in a story—you have a reason to add an example. If it hits the brief, keep the zero-shot version. It is shorter to maintain, and you are not spending context on samples you do not need.
When should you add examples?
Add examples when the instruction is clear and the result still drifts in ways you can point to: the wrong heading pattern, a voice that does not match the rest of the book, or a workbook page that keeps turning into an essay. That is the situation Google is describing when it says few-shot examples help regulate formatting, phrasing and scope.
Anthropic calls examples one of the most reliable ways to steer format, tone and structure, and it suggests three to five examples that are relevant, diverse and clearly marked. Google also warns that too many examples can cause the model to overfit to the samples, and it stresses consistent formatting—including XML tags, whitespace and separators—because a main job of the examples is to show the response format.
A weak few-shot prompt dumps two long chapters into the request and says “write the next one like these.” A stronger version shows only the pattern you care about, and it varies the examples enough that the model does not copy an incidental detail.
Here is a compact before-and-after for chapter openings. The zero-shot brief above is the starting point. If openings still wander, add two short samples:
Write the opening for the next chapter using the same pattern as the
examples. Keep the length to 200 to 250 words. Use only the notes for
this chapter. Do not copy the examples' topic or sentences.
Example 1
Notes: The reader is a new freelance designer who underprices because
they compare themselves with students, not with paid work.
Job: help them set a floor rate before the next inquiry.
Opening:
You are not imagining it: it is hard to name a rate when your last
"project" was a favor. This chapter will help you set a floor you can
say out loud before the next inquiry arrives. You will write that floor,
test it on one unpaid request, and keep a one-line reply ready.
Example 2
Notes: The reader avoids sending invoices on time because the work
felt informal.
Job: help them send a same-week invoice without writing a long apology.
Opening:
If the project felt friendly, the invoice can feel abrupt. This chapter
will help you send a same-week invoice without a long apology. You will
pick a send day, use a short three-line message, and record the date so
you can follow up once.
Now write the opening for this chapter.
Notes: The reader says yes to extra revisions because they are afraid
of seeming difficult.
Job: help them state a revision limit before work starts.
The examples teach the pattern: name the situation, state the job, preview the steps. They do not teach the new chapter’s content. If you need a fuller treatment of how to lift one of your own chapters into a reusable sample, use the dedicated few-shot chapter guide.
How should you format the examples?
OpenAI’s guide typically places examples inside a developer or instructions message, and it shows them wrapped in simple XML-style tags, such as a product review followed by the desired label. Anthropic recommends wrapping examples in tags so the model can distinguish them from instructions, and using consistent, descriptive tag names. Google says the structure of the examples should match, especially punctuation, whitespace and any XML tags you use.
For an eBook, a reliable pattern is:
- State the standing rules once.
- Show two or three labeled examples in the same shape.
- Give the new notes and the request last.
If your examples include an odd heading, a joke, or a fact that is only true in that chapter, expect some of that to leak. Diverse examples help. One teaching chapter and one exercise chapter, both following the same skeleton, teach the skeleton more cleanly than three nearly identical openings.
You should also keep the examples honest. If you want short openings, do not include a 600-word sample. If you want the model to stay inside supplied notes, do not show an example that invents a statistic. The model will pick up the pattern you demonstrate, including the accidental ones.
A worked comparison for a workbook page
Suppose you are writing a budgeting workbook for freelancers. You want each chapter to end with a five-minute page: a one-sentence instruction, three lines to fill in, and a backup question. A clear zero-shot prompt may still produce a quiz, a long reflection, or a page with no space to write.
In that case, show two completed pages and then ask for the third. Keep the samples short. Vary the topic so the model copies the page shape rather than the subject. Then inspect the new page against a checklist: Can the reader finish it in five minutes? Are the lines actually fillable? Did any advice appear that was not in the chapter notes?
If the first few-shot result overfits—say every page now mentions “client calls” because both samples did—replace one sample with a different situation, or cut back to one-shot and strengthen the format instructions. Google’s overfitting warning is useful here: more examples are not automatically better.
How does this fit a longer book project?
For a whole eBook, you rarely want to paste three full chapters into every request. Keep standing rules and a few short examples in a reusable instruction, and put the current chapter’s notes in the request. That is the same split OpenAI describes between instructions and input, and it is the reason a template is often the right next step after you have a working few-shot prompt. Our guide to writing a prompt template covers that move.
In Interactive eBook Builder, Spec Builder, Drafter and Polisher already split planning, drafting and editorial refinement into separate stages, which keeps each prompt closer to one job. You can attach approved examples to the stage that needs them—openings for Drafter, for instance—without asking one prompt to plan, write and polish the whole manuscript. SEO BOT can later publish supporting articles to the Auto-Blogger blog, each linking back to your eBook, site and brand. As always, review the work before it goes out.
FAQ
How many examples should I use?
There is no single official number that fits every task. Anthropic suggests three to five relevant, diverse examples as a useful range. Google says models can often pick up a pattern from a few examples, and that you may need to experiment, while too many can overfit. Start with one well-chosen sample if you have it, then add a second or third when a specific miss remains.
Do reasoning models still need few-shot prompts?
Not always. OpenAI’s reasoning-model guidance says those models often do not need few-shot examples, and it recommends trying a prompt without them first. If you add examples, they should align closely with the instructions. If you use a reasoning model and the zero-shot brief already meets your requirements, you can leave the examples out.
Can I use examples from someone else’s book?
Use examples you have the right to use, preferably from your own draft. You want the model to copy your structure and voice, not another author’s sentences. Short samples you have approved are more useful than long passages from a book you admire.
Why did few-shot make the answer worse?
Usually the examples are fighting the instructions, repeating the same pattern, or teaching an incidental detail. Check that each example would pass your own requirements, that the formatting matches, and that the new request is clearly marked as the thing to produce now. Then test again with one example removed.
Before you hit send
- Did you try a clear zero-shot brief before adding examples?
- If you added examples, do they show the pattern you want and nothing you do not want copied?
- Are the examples formatted the same way, and clearly separated from the new request?
- Do the examples stay inside notes, length and tone rules you would accept in a finished chapter?
- Are you judging the result against your requirements, not against whether it resembles the samples at a glance?
- Have you tested the prompt on a second set of notes before reusing it across the book?