Few-Shot Prompting: Show Two Examples Instead of Describing What You Want

You've typed "write in a friendly, professional tone" into ChatGPT and gotten back something that reads like a bank statement wearing a cardigan. Then you rewrote it. Then you did the same thing the next day.

The instruction wasn't wrong. It was just too abstract to be useful. "Friendly, professional" describes maybe four hundred different writing styles, and the model picked the blandest one because that's the safe average. Few-shot prompting fixes this by skipping the description entirely: you paste in two or three things you've already written, and let the AI copy the pattern.

It is the single highest-return technique we teach, and it takes about ninety seconds to set up the first time.

What few-shot prompting actually is

"Shot" is a term of art that means "example." Zero-shot is what most people do — you describe the job and hope. Few-shot is when you show the model two to four completed examples before asking for a new one.

That's the whole idea. No special syntax, no settings to change, nothing to install. You paste examples into the same chat box you already use.

The reason it works so well is that an example carries decisions you'd never think to write down. How long your sentences run. Whether you use the customer's first name. Whether you sign off with "Thanks" or "Best" or nothing at all. Whether you apologize once or twice. You could spend twenty minutes listing those rules and still miss half of them. Two real emails carry all of it at once.

Building one from work you've already done

The material is already sitting in your sent folder. That's the part people miss — you're not writing examples, you're finding them.

Pick a job you do repeatedly. Then find two or three times you did it well — a reply that got a good response, a post that actually performed, a summary your team didn't have to ask questions about. Paste them in with consistent labels, then hand over the new item.

I write follow-up emails after quoting a job. Here are two I've
sent that got replies, so you can match how I write:

---
Situation: Quoted a kitchen backsplash, no response for 5 days.
My email: "Hi Marcus - just circling back on the backsplash
quote. No rush at all, but the tile we spec'd has a 3-week lead
time right now, so I wanted to flag that before it stretches.
Happy to look at a cheaper option if the number was high.
- Dave"
---
Situation: Quoted a full bathroom, customer said "let me talk to
my wife."
My email: "Hi Jen - no pressure, just checking in. If it helps
the conversation, I can break the quote into phase 1 (plumbing
and tile) and phase 2 (fixtures) so it's easier to stage the
cost. Let me know either way and I'll close the file.
- Dave"
---

Now write one in the same voice:
Situation: [DESCRIBE THE NEW SITUATION]

Look at what those two examples smuggled in without a single rule being stated. Short. First name, no "Dear." A specific reason to reply that isn't pressure. An offer to make it cheaper or smaller. A one-word sign-off. Try describing that in a paragraph — it takes longer and works worse.

If your business has a voice you want locked in across everything, not just one email type, that's what the brand voice codifier is for. It turns your best existing writing into a reusable description you can paste at the top of any prompt.

The four rules that make examples work

Most few-shot prompts that underperform break one of these.

• Two to four examples, not one and not ten. One example can be a fluke and the model will over-copy its specifics. Past four, you're mostly burning time for no gain on everyday tasks.
• Use real material, not samples you wrote for the occasion. Invented examples drift toward the same generic average you're trying to escape. Your sent folder is more honest than you are.
• Keep them consistent with each other. If one example is chatty and one is formal, the AI splits the difference and you get something that sounds like neither. Pick a lane.
• Label the new item exactly like the examples. If you wrote "Situation:" and "My email:", use those same labels for the new request. The model is completing a pattern; give it a clean one to complete.

One more that isn't a rule so much as an unfair advantage: include a hard case. An example of how you handled an angry customer teaches the model more than three examples of you handling happy ones. Edge cases carry your judgment, and judgment is the thing that's hardest to describe.

Where it earns the most, fast

Few-shot is not equally useful everywhere. It pays off hardest on jobs that are high-volume, style-sensitive, and already have a paper trail.

• Review replies. You've written dozens; two of them are your best; you'll never write a description of your review voice that beats just pasting those two. Our review response writer is built around exactly this pattern.
• Support responses. Show the model three of your existing canned answers and it will write new ones that slot into the same set — which is how a library of support macros grows without sounding like it was assembled by committee.
• Sorting and tagging. Show how you label five incoming inquiries ("billing," "warranty," "just browsing") and the AI will tag the rest of the inbox the same way. This is the quiet workhorse use, and it's what makes triaging an inbox actually work.
• Structured summaries. One perfect meeting summary as an example beats any amount of instruction about what the summary should contain.

Notice the pattern: these are all jobs where you already know what good looks like but can't easily articulate it. That gap is precisely where examples beat descriptions.

Few-shot and the four-part brief work together

These aren't competing techniques. Examples handle style and shape; the brief handles everything examples can't carry.

The role-context-task-format framework covers who the AI is, what it needs to know about your situation, what job to do, and what the output should look like. Few-shot slots into the format part, and does that job better than words do. In practice a strong prompt for a recurring task is a short brief plus two examples — maybe six lines of instruction and two pasted emails.

That combination is also what separates a throwaway prompt from something you save and reuse, which we get into in master prompt vs. regular prompt.

Where this breaks

Examples steer style and structure hard. They do not check facts, and this catches people out.

If one of your examples mentions a 10% discount, a two-week turnaround, or a warranty term, the model may cheerfully carry that detail into new drafts where it isn't true. It's completing a pattern, and your promises are part of the pattern it sees. Strip specific numbers and commitments out of your examples where you can, and read every draft for names, dates, prices, and anything you'd be held to.

The other limit is drift. If your examples are three years old and your business has changed, few-shot will faithfully reproduce a voice you've outgrown. Refresh them when the work changes.

And for anything touching money, employment, or contracts, matching your voice does not make the content correct. Your accountant and your lawyer still exist for a reason.

Try it on one job this week

Open your sent folder, find two things you wrote that worked, and build one few-shot prompt around them. Save it in a note. The setup cost is once; every use after that is free.

If you want the technique taught properly with exercises, it's covered in the few-shot lesson inside our free prompt engineering course, and the prompt improver will show you what your current prompt is missing on your own work.