Understand AI Marketing

How Do Examples Help an AI Tool Understand What You Want?

Understand how examples can show an AI tool the form, detail, tone or boundary you want, and how to prevent unwanted patterns carrying into the output.

2 September 2026By Michael Sweenie7 min read

An example can show a pattern that written instructions leave open, such as structure, detail, tone or a decision boundary.

Use a relevant, approved example with a clear instruction. Label what to follow and what not to copy. Examples can pass on unwanted wording, facts, bias or formatting, so review remains essential.

What is an example in an AI prompt?

An example is a sample input, output or input-and-output pair placed in the prompt for the tool to use as a pattern.

A request with no example is often called a zero-shot prompt. A request with a small number of examples is often called a few-shot prompt. You do not need those terms to use the idea.

The roles are more important:

  • Instruction: what the tool should do now.
  • Example: what one acceptable pattern looks like.
  • Source material: the facts, notes or copy the tool should examine.

An example is one kind of context, but it is not automatically factual evidence. If you show the tool an approved article introduction, its shape may be useful while its subject and claims are irrelevant to the new task.

What Is a Prompt in AI Marketing? explains how to make the task, audience and output clear. An example adds a demonstration when words alone still leave an important choice open.

Four things an example can make visible

SignalWhat an example can showWhat you should still state
FormLabels, order, paragraph shape or table fieldsWhich parts of the deliverable are required
DetailApproximate depth and amount of explanationAny exact length or coverage requirement
TonePlain vocabulary, sentence style and claim strengthThe audience, purpose and voice boundaries
Decision boundaryA positive case, counterexample or treatment of an unknownThe rule and what to do with an ambiguous case

These four signals are a practical way to annotate an example, not a universal prompting formula.

What current provider guidance shows

Google's prompt design guide, last updated 10 June 2026, says examples can show patterns for formatting, phrasing and scope. It also says the number of examples may need experimentation and warns that too many can make a response fit the examples too closely.

Google documents a small Gemini 2.5 Flash comparison. With no examples, the recorded response selects the longer of two explanations. After two earlier examples show a preference for concise answers, the recorded response selects the shorter one.

That is Google's documented demonstration, not a test performed for this article. It shows that examples can influence one response pattern in one named task. It does not prove that an example will improve every marketing output or behave identically in another tool.

Anthropic's current prompting guidance similarly describes examples as a way to guide format, tone and structure. It recommends examples that reflect the real use case and vary enough to reduce unintended patterns. That is provider-specific guidance, not a universal performance guarantee.

An annotated marketing example

ILLUSTRATIVE AND UNTESTED: The paragraph and tasks below are fictional. They are not client material, employer content, a model test or a measured result.

Imagine the current instruction is:

Draft a short introduction explaining why a claim-level source register helps a small B2B marketer. Use clear UK English and a measured tone.

You could add this fictional example:

A source register gives each research claim a place to live. It records the source, supporting passage and checking status before the claim enters a draft. That makes uncertainty visible while the marketer keeps the final decision.

Before using it for a different subject, annotate the pattern.

FollowDo not copy
Start with a plain definitionThe source-register subject
Use three short sentencesThe phrase about a place to live
Use clear UK EnglishExample-specific facts or claims
Keep the claim strength measuredThree sentences if the new task needs another shape
End with the marketer's rolePersonal, client, customer or employer details

For a new task, make the relationship explicit:

Draft a short introduction explaining why a review checklist helps a small B2B marketer check AI-assisted copy. Follow the example's definition-first order, short sentences, plain UK English, measured claims and final human-control point. Do not reuse its subject, facts or distinctive phrasing.

This comparison shows the extra information supplied by the annotated example. It does not claim that an AI tool produced a better response.

How examples can go wrong

An AI tool can pick up a feature you did not intend to teach.

RiskWhat may happenPractical response
Accidental wordingA memorable phrase reappearsName the general quality and prohibit phrase copying
Irrelevant factsDetails from the example enter the new draftStrip unnecessary facts and separate source material
Narrow patternOne layout looks like the only valid answerMark flexible elements or add a genuinely different approved example
Conflicting signalsThe instruction and example request different formatsCorrect the conflict before running the task
Weak sourceErrors, hype or bias are repeatedReview the example before asking AI to follow it

Do not use a competitor's copy or another creator's recognisable voice as a shortcut. Use your own approved work or an authorised, generalised example and focus on observable qualities.

How many examples should you use?

There is no universal number. Start with one short example when it resolves one real ambiguity. Add another only when it demonstrates a necessary variation, difficult case or boundary.

More examples create more material to inspect and more patterns the tool may carry across. Provider recommendations also differ and can change with models. Test the smallest useful set on your actual task.

If you need to decide which background belongs in the request at all, read What Does Context Mean in AI Marketing?.

Check the output, not just the prompt

After the AI responds, ask:

  1. Did it complete the current task rather than repeat the example's subject?
  2. Which intended qualities did it follow?
  3. Which words, facts or quirks did it copy accidentally?
  4. Did it respect the new audience, evidence and constraints?
  5. Are the claims accurate and supported independently of the example?

If the wrong feature carried across, improve the annotation or replace the example. Do not assume that a fluent resemblance means the output is suitable.

Your next step

Choose one short example you own or are authorised to use. Mark three qualities under Follow and three details under Do not copy. Use it for one small task, then review the response for both intended and accidental patterns.

Further Reading

You Might Still Be Wondering...

Frequently asked questions

Back to Blogs