Show, don't tell: examples beat descriptions

The one idea

Describing what you want is hard. Showing it is easy.

"Write product update notes in our style — friendly but not jokey, short sentences, benefit first, feature second…"

You will go back and forth four times. Instead:

"Here are two product updates we published before: [example 1] [example 2]. Write the next one, matching their style, for this change: [your notes]."

The model is a pattern machine — examples are patterns. This is the single highest-leverage trick in prompting, and practitioners call it few-shot prompting.

Where it shines at work

• House style — emails, updates, reports that sound like your team
• Repeating formats — meeting notes, status updates, CV screening summaries: show one filled-in example, then feed it new raw material each time
• Tricky judgment calls — show three customer messages you labeled urgent and three you labeled routine; the model picks up the boundary far better than from your description of it

How to pick examples

1. Use real, good examples — the model will copy whatever you show, flaws included
2. Two or three is usually enough; one works for style copying
3. If output keeps getting something wrong, add an example that gets that exact thing right

Try it now

Find a format you produce regularly. Paste your best past example, then the new raw input, and ask for "the next one in exactly this format and style".

Your win today

You now know the answer to "the AI never gets our tone right" — stop describing the tone, paste two examples of it.