The most useful line in a prompt is often the one that says "don't"

December 18, 2022 was a Sunday, and as far as we can tell, the AI world took it off. ChatGPT was a little over two weeks old (it launched November 30), and we couldn't find a launch, paper or announcement dated to this day worth writing up. So no fake news here. A lesson instead.

What people kept getting back

If you've used a chatbot much, you probably know the shape of a default answer. A friendly opening line restating your question. A short preamble. The actual content, usually as a list. A closing paragraph that summarizes the list you just read, maybe with a reminder to consult a professional.

None of that is wrong, exactly. It's just padding you didn't ask for, and it comes back every time because you never said not to.

People spend a lot of effort describing what they want. Fewer spend one line on what they don't. That line is cheap, and it often does more work than the rest of the prompt.

Before and after

Before:
Give me a packing list for a three-day work trip.

After:
Give me a packing list for a three-day work trip, carry-on only.
Leave out toiletries (I have a kit) and anything obvious like
"phone" or "wallet." No intro, no closing tips. Just the list,
grouped by clothes and work gear.

The before version gets you a list padded with toothbrush, charger, passport, and a cheerful note to check the weather. The after version gets you the eight or nine things you might actually forget. Same model, same task. The difference is that the second prompt drew a border.

Make the "don't" specific

This is where it goes wrong for people. Vague bans don't do much.

"Don't be generic" doesn't tell the model what generic means to you. "Don't open with a definition of the topic" does. "Keep it simple" is a wish. "No words a 12-year-old wouldn't know" is a rule it can follow.

It also helps to pair the ban with the replacement. "No intro, start with step one." "Skip the history, begin with how it works today." The model knows what to put in the space you cleared.

And keep the list short. A prompt with fifteen "never" rules stacked at the bottom tends to get an answer that reads nervous, like the model is tiptoeing. Pick the two or three things that annoy you most and name those.