Show, don't tell: use examples to get the exact format you want

The fastest way to get the model to match your format is to show it one thing that already does.

What you'll have when you're done

A simple habit you can use in any chat window today: paste one example of what good looks like, tell the model to match it, and watch the format land correctly on the first try. No more describing your preferred style in paragraphs and still getting something that looks wrong. The example is the instruction.

The "describe it three times and still not get it" problem

You've done this. You write: "Give me meeting notes in a clean, punchy, scannable format." You get back a wall. You add: "Actually make it shorter, with clear headers." Still not right. You add: "I want it to look more like an executive summary." The third version is closer, maybe, but it's still not how you'd write it.

Describing format is genuinely hard. "Punchy" means something different to a model trained on every type of writing that exists. "Clean" is even worse. An adjective is interpretable a dozen ways. You are trying to transmit a very specific thing, and the transmission medium (description) is lossy.

The highest-ROI fix for format problems is to stop describing and start showing. Paste one example of what good looks like. The model reads it, extracts the shape, and builds the new output to match.

Per Anthropic's prompting guidance, examples are one of the most reliable ways to steer output format and tone. A single well-chosen example often replaces three or four sentences of description, because the example communicates format, length, tone, and structure all at once. It is dense, unambiguous signal. Description is not.

This technique shows up as "few-shot prompting" in AI research, but you do not need to know or care about that term. The habit is simpler than the name: show, don't tell.

What you need first

• A Claude or ChatGPT account. Plain chat, no terminal, no setup.
• One real output you've been unhappy with. Meeting notes, emails, a report: pick the one where you keep ending up in revision cycles.
• One good example from your past. A set of notes you liked, an email you wrote yourself, last month's report. Something that already looks the way you want.

No technical background needed. You are pasting text and adding one sentence.

Step by step

The core move

Open your chat window. Before you give the model a task, do this:

1. Write: "Here is an example of what good looks like:"
2. Paste your example.
3. Then give the task: "Now do [today's thing] in exactly this shape."

That is it. The model is a pattern-matcher at its core. It does not work the way a calculator does, executing rules you type. It predicts the next most-useful output given everything you showed it. When you show it a complete example of the output you want, the prediction lands on something that looks like the example, because the example is the densest signal in your prompt.

The anatomy of a prompt has four parts (task, context, format, constraints). An example compresses all four into one paste. You do not need to write out the format rules if you can paste a thing that already follows them.

Three situations where this works immediately

1. Meeting notes in your shape

You have a format you like. Maybe it is three sections: decisions, action items, open questions. Maybe it is shorter than that. The model does not know.

What to do: paste one past set of notes you were happy with, say "Format today's notes in exactly this shape," then paste the transcript or your rough summary. The model reads the example, sees the section structure, the length, the level of detail, and it builds the new notes to match.

2. Emails in your voice

Voice is the hardest thing to describe. "Direct but not cold." "Casual but not sloppy." These are real instructions that land differently for everyone.

What to do: paste two short emails you actually wrote. Do not make them up. Then say: "Draft a reply to the message below in this same voice and length." Two real examples is often enough because they share a pattern: your sentence rhythm, your sign-off, whether you open with context or cut straight to the ask. The model reads both and extracts what they have in common.

3. Monthly update in your structure

You send the same report every month. Same sections, same table, different numbers.

What to do: paste last month's update. Say: "Do this month's in the same sections and table format, swapping in the new numbers." Then paste this month's raw data or notes. The model carries the shell forward. You get a draft that is already in the right shape, with the right sections, at the right length. You are editing numbers, not rebuilding the document.

How to do this well

Use a real example, not a made-up one. If you write a fake example just to have one, you are teaching a fake pattern. The model will reproduce the fake, not your actual style. Pull from something you actually wrote or approved. The example is only as good as the original.

Label it clearly. "Here is an example of what good looks like:" tells the model explicitly that this is a reference to match, not content to respond to. Without the label, the model sometimes treats the pasted example as input to analyze or comment on, not a template to follow.

One to three examples is the right range. One is usually enough for format and length. Two is useful for voice, because the model can find what they share. Three is the ceiling. More than three can over-constrain: the model starts averaging across all the examples and you lose the specific edge you were trying to preserve.

You can stack the example with context and constraints. Paste the example, then add: "The audience is my leadership team. Keep it under 200 words. Use the same sections but skip the 'risks' row this month because we do not have an update." The example handles the shape; the rest of your prompt handles the specifics of this instance.

How you'll know it's working

The format lands on the first pass. You are not in a revision loop over headers, length, or tone. You might edit for accuracy, for a word choice here or there, but the structure is already what you wanted.

You stop writing format descriptions. You no longer need "clean, punchy, scannable." The example handles all of that in one paste.

You start saving examples for reuse. Once this habit clicks, you will naturally start saving good outputs so you can use them as examples again. That is the compounding version of this habit: a small library of reference outputs you paste whenever you need that shape.

When it breaks

The example is sloppy or off-target. A mediocre example teaches a mediocre pattern. If you paste a set of notes that you were only halfway happy with, you will get output that is only halfway what you wanted. The example has to be something you actually want to replicate. Garbage in, garbage out applies here more directly than almost anywhere else in prompting.

The example is too long. If you paste a 1,500-word report as your example, the model may treat length as part of the pattern and produce a 1,500-word version of your three-bullet update. When you want a short output, show a short example. Add a length constraint if there is any ambiguity: "Match this format but keep the total under 150 words."

The format is still wrong after a clean example. Most of the time this means the model is treating the example as content to reply to, not a template to match. Add the label ("Here is an example of what good looks like:") and re-run. If it is still off, check whether the briefing method itself needs work: why AI gave you a bad answer walks through the common culprits.

Level up

Once you are using examples to lock format, the next move is making context persistent so you stop rebuilding context in every session. The full briefing method covers how to combine the example technique with role, context, task, and constraints into a complete, reusable prompt structure.

If you want to understand why showing an example works at the model level, how AI works in plain English goes into the prediction mechanics without the jargon. Short version: the model is always predicting from patterns. Hand it a strong pattern, get a strong prediction back.

Save one output you were happy with this week. The next time you ask for the same type of thing, paste it first. One sentence: "Here is an example of what good looks like." The revision cycle you have been in will be shorter. Probably much shorter.