Most people write long instructions when they want an AI to match their voice. There is a faster approach, and it works better.
Before you ask Claude or ChatGPT to draft this quarter’s board memo, paste in last quarter’s, the one you were actually happy with, and say “like this, for Q3.” That single move beats any set of written instructions you’ll spend twenty minutes crafting.
Why examples outperform instructions
Instructions describe what you want. An example shows it. You can type “two pages, confident but not salesy, lead with the number,” and the model has to guess what each of those phrases means to you specifically. Hand it a finished memo you’d put your name on, and it can see the tone, the length, how you open, where you bury the bad news, and everything you leave out.
Both major AI providers say the same. Anthropic’s prompting guide for Claude calls examples one of the most reliable ways to steer tone and structure. OpenAI’s guide for ChatGPT categorizes it under few-shot learning, the research term for showing a handful of finished outputs instead of describing what you want.
Two things that determine whether it works
First, pick the example you’re proud of, not the average one. The model copies what you give it, so a mediocre example produces a mediocre draft, just faster.
Second, if you don’t have a good example yet, you’re one round away from building one. Take the first draft the model gives you, edit it the way you normally would, and save that edited version. That becomes your example for next time. The second ask is always better than the first. By the fifth, it barely needs direction.
Where this applies
The technique works on anything you’ve done well at least once: customer summaries, deck outlines, weekly updates, replies to the board member who asks the same question every quarter. Show it instead of describing it, and what comes back reads like you wrote it on a good day.
