Show, don't tell

If you can only learn one technique from this folder, learn this one.

Models are dramatically better at copying a pattern than at following an instruction about a pattern. Describing the shape of the output you want — even in detail — produces drift. Pasting one good example of the shape kills the drift instantly. This is called few-shot prompting and it's the closest thing AI has to a free lunch.

The numbers

In practice, on real knowledge work:

• Zero examples → output shape is whatever the model felt like.
• One example → output shape locks to that example most of the time.
• Three examples → output shape locks reliably, and the model picks up subtle conventions in the examples (capitalization rules, ordering, edge-case handling) without being told.
• Five+ examples → marginal gains. Save your tokens.

For most prompts you'll write, one example is the sweet spot. Two if the format is non-obvious. Three if you've tried and the model keeps drifting.

The "match this voice and structure" pattern

The most useful shape of an example-based prompt:

You are [role].

Task: [what you want done].

Here is an example of the output style and structure I want. Match
this voice and structure exactly — substitute the content, keep the shape:

---
[Paste a real previous output here — an old report, an old email, an
old executive summary. The actual thing, not a cleaned-up version.]
---

Now do the same for the attached [file]. Output: [filename + format].

That's the whole pattern. Paste in a previous month's distributor scorecard before asking for this month's. Paste in an email you actually sent before asking for the next one in the thread. Paste in a slide deck outline before asking for one on a different topic.

The model will reproduce the structure — section headings, the order things appear, the rhythm of bullets vs prose, whether you start with a number or a sentence — far more reliably than if you described all of that in words.

When zero-shot is fine

Not every prompt needs an example. Skip the example when:

• The task is well-known and standardized (summarize a contract, draft a polite decline email, transcribe meeting notes). The model has seen enough of these in training that its default shape is fine.
• You don't have an example handy. "Best two-minute prompt now" beats "perfect prompt in twenty minutes." Try zero-shot first, see if you can live with the output.
• The cost of being wrong is low — exploratory work, brainstorming, throwaway drafts.

Use examples when:

• The format is specific to your company (your distributor scorecard, your QBR template, your CFO's preferred summary length and tone).
• The format is novel (you're inventing a new report shape this month).
• You've tried zero-shot and the model keeps producing the wrong shape.
• The output is going to a customer or executive and consistency matters more than creativity.

Where to find your examples

The best example is a piece of work you've already done and were happy with. Specifically:

• The last well-received version of the same deliverable (last month's scorecard, last quarter's recap, last week's email).
• A peer's version of the deliverable that has the voice you want.
• A best-practice version from a vendor — e.g., a vendor's own published case study, when you're drafting one.

A second-best example is one you sketch yourself in 60 seconds — even a structurally-correct, intentionally-thin version. The model copies the structure far more than the content, so a thin-but-correctly-shaped example beats a verbose-but-vaguely-described instruction.

A small but important detail: where to put the example

Put the example before the new input the model is supposed to transform.

Good: "Here's an example. [example] Now do this for the attached file."

Bad: "Do this for the attached file. Here's what good looks like: [example]."

The model uses what came earlier in the prompt to decide what to do with what comes later. The example needs to be in place before the task lands.

Common mistakes

• Describing the example instead of pasting it. "Format it like a typical executive summary" is not an example. The actual paragraph of a typical executive summary is the example.
• Pasting a perfectly polished example. The cleaner the example, the less the model picks up on real-world conventions like footnotes, caveats, or "as of [date]" lines. A real example beats an idealized one.
• Too many examples. Three rarely beats two. Five rarely beats three. Use what's necessary, not what's available.
• Pasting the wrong kind of example. If you want the model to produce an Excel scorecard, an example email won't help. The example has to be the shape of the output, not just from the same domain.

Where to go next

• The four-part prompt — the structure that holds your example inside it.
• Make it a critic — what to do after the first draft comes back; ask the model whether it actually matched your example.
• Decompose and verify — for longer outputs, you may need an example per section rather than one example for the whole deliverable.