Using examples and output formats to control results

0:00 Examples and output formats

What does witty mean? To you, it might mean dry, understated and specific. To the AI, it might mean puns and three exclamation marks. That gap is why so many people describe their brand voice perfectly and still get output that sounds nothing like them. In this lecture you'll learn the fastest fix in all of prompt engineering: examples. And its partner: output formats, which make results easier to review and paste into your tools. You'll see a UK boutique's subject lines, an example bank for a skincare brand, and watch me turn adjectives into examples live.

0:42 Why examples win

Why do examples work so well? Language models are exceptional pattern followers. When you include examples, a technique often called few shot prompting, you give the model a concrete pattern to match. Friendly, confident and slightly witty means different things to different people. But here are two captions in our voice, write three more, leaves very little room for misunderstanding. The model can see your sentence length, your punctuation habits, how you open, how you close, how much you joke. Here's the key idea. Adjectives describe a voice. Examples demonstrate it. And demonstration wins almost every time.

1:24 The music analogy

Think of it like asking a band to play something in the style of your favorite artist. You could describe it: upbeat but a bit moody, with a strong chorus. Or you could just play them two tracks. Two tracks work better than one, because the band can hear what's consistent across both, the style, and what's specific to each, the song. That's why two or three varied examples beat a single example. And you'd add: same style, new song. Don't copy the melody. In prompts, that's: match the tone and structure, not the topic or specific phrases.

2:07 Five rules for examples

Five rules for using examples well. One, use your best real work: posts that performed, emails that got replies. Two, show variety within the pattern: two or three examples on different topics with the same voice teach the model what to keep constant. Three, label them: example one, Reel caption, product launch. Four, say what to copy and what not to: match tone and structure, not topics or phrases, and don't reuse any phrase of four or more words. And five, use counter examples sparingly: here's a caption we dislike because it's too salesy. That sharpens the boundary without flooding the prompt with bad patterns.

2:53 Output formats

Now formats. Asking for a specific structure makes output easier to review, compare and move into other tools. Numbered options with a rationale: ten hooks, each followed by the angle in brackets. Tables: a two week content calendar with columns for date, platform, format, hook, key message and call to action. Scripts with timing: time, visual, voiceover. CSV with headers for pasting into a spreadsheet or scheduler. Or a fixed template: hook, problem, solution, proof, call to action, one line each. One warning: a beautiful table of weak ideas is still weak. Format supports thinking. It doesn't replace context.

3:36 Example 1: the summer sale subject lines

A simple example. A UK online fashion boutique wants subject lines for a summer sale. They paste three past subject lines with high open rates: the linen edit you asked for, finally. Your Saturday outfit, sorted. We restocked the dress, yes, that one. Then: match their voice, conversational, specific, a little knowing, no shouting. Avoid all caps, excessive emojis, and the word sale in the first three words. Write twelve subject lines for the summer sale, up to forty percent off selected lines, ending Sunday. Format: a table with subject line, preview text under ninety characters, and angle. The result is consistent, on brand and easy to shortlist, and the team tests the top options with their email platform's A B feature.

4:29 Example 2: the skincare example bank

Now a business scenario, with illustrative details. A skincare brand in Dubai builds an example bank: top Reel captions, best subject lines, sales emails that got replies, and a short counter examples section. They store it in their brand project. A month later, outputs start including the phrase clinically proven to reduce wrinkles. Nobody asked for it. The culprit? An old caption in the example bank, written before legal withdrew that claim. The model was doing exactly what it was told: matching the examples. They remove the old caption, add a note to review the bank every quarter, and the claim disappears. Examples are powerful in both directions.

5:16 Watch me do it

Let me show you the difference live. First, adjectives only: write five Instagram captions for a candle brand in a warm, witty, cozy tone. Here's the result: lots of puns about wick, and three exclamation marks. Now I attach the example bank and say: using the Reel captions section, write five captions for our new autumn candle. Match tone, sentence length and structure, not topics or phrases. Don't reuse any phrase of four or more words from the examples. Format: a table with caption, angle, and why it fits the examples. The captions are short, dry and specific, exactly like the brand's top performers. And I quickly scan for copied phrases. None. That's the power of show, don't tell.

6:08 Formats that feed your tools

One more format tip for anyone who moves AI output into other tools. When content is headed for a spreadsheet, a scheduler or a CRM import, ask for a machine friendly format with named headers. Return only CSV with headers date, platform, hook, caption, call to action and hashtags, and quote any field that contains commas. Then test one small batch before you scale up. Paste it into a sheet and check every column landed where it should. Captions with commas are the classic problem, and one misaligned row can shift a whole content calendar. Once the format works, save that exact instruction in your prompt library, so it works the same way every time, for everyone on the team.

7:00 Common mistakes

Common mistakes. Copying examples too closely. If outputs repeat your example phrases, add: don't reuse phrases from the examples. Examples that break your own rules, like an old claim or a missing disclosure. The model will copy the problem. Too many examples. Beyond a handful, returns diminish and prompts get long, so choose your best. And format without substance: a beautifully formatted table can hide weak thinking. Always ask whether the ideas in the table are actually good.

7:34 Recap and try this now

Let's recap. Examples transfer voice better than adjectives, because models are pattern followers. Use your best real work, show variety, label your examples, say what to copy and what not to, and keep counter examples sparing. Ask for the output format you actually need, from tables to CSV. And keep your example bank clean, because the model will copy whatever's in it. Here's your try this now. Spend twenty minutes building an example bank from the template in the lesson text, with three to five top performers per content type. Store it in your project, and run your next content request against it.