The chatbot ran long in week two. It still does, unless you give it a number.

You ask a simple question. What temperature should chicken be cooked to? The answer comes back in five paragraphs, with a heading, a note about food safety culture, a reminder to consult a professional, and somewhere in the middle, the number you wanted.

That's a made-up example, but it's the kind of answer OpenAI itself warned about in ChatGPT's first weeks. On Sunday, December 11, 2022, I can't find anything dated to that day worth writing up. No launch, no paper, no announcement I could verify. The record is quiet. So instead of inventing a headline, here's the lesson that week kept teaching anyone who used the thing: tell it how long.

OpenAI said so on day one

This wasn't a secret. The launch post OpenAI published on November 30 had a limitations section, and one line in it is unusually frank. It said the model "is often excessively verbose and overuses certain phrases, such as restating that it’s a language model trained by OpenAI." The reason it gave: trainers preferred longer answers that looked more comprehensive, plus what the post called well-known over-optimization issues.

I like that explanation because it's so human. People rating answers saw a long one and thought "thorough." The model learned that long gets rewarded. Then everyone else had to scroll.

Why "be brief" doesn't work

The natural fix is to say "keep it short" or "be concise." It helps a little. But short means different things to different people, and the model fills that gap with its own idea, which tends to drift long.

A number removes the guess. "Under 60 words." "Three sentences." "One line per step, no more than five steps." Now there's something to measure against, both for the model and for you.

A few things tend to make the limit stick:

• Put the limit near the end of the prompt, close to where the model starts writing.
• Give a reason. "It's going in a text message" explains the limit and the tone in one go.
• Prefer sentences or lines over exact words. Models count words loosely. They're better at "three sentences."
• If it overshoots, don't just say "shorter." Say "cut this to half its length and keep the second point."

That last one matters more than it sounds. Vague complaints get vague edits.