09of tenWorking Notes on Prompting · What the Tool Is Actually Doing

The Length and Concision Trap

A working note on what "be concise" and "in 500 words" actually do to your output.

"Be concise." "Keep it brief." "In exactly 500 words." "One paragraph." These feel like precise controls, dials you set on the output. They're better understood as rough suggestions the model interprets by feel, and they often cost you more than the length they save.

The common move

The familiar prompt Explain this concept concisely, in exactly 500 words.

You've named a tone (concise) and a number (500), so it feels like you've set firm constraints. The question worth asking: can the model actually do either of those things?

What you think it does, and what it actually does

What you think it does
  • Counts the words. "500 words" feels like a number the model hits exactly.
  • Improves the writing. "Concise" feels like it can only make the answer better.
  • Is a precision tool. A length cap feels like fine control.
What it actually does
  • Estimates, and misses. The model doesn't count as it writes; it approximates. You ask for 500 and get 340, or 700.
  • Pays in substance. "Concise" often strips the reasoning and nuance that made the answer worth having. You don't see what got cut.
  • Is a blunt one. The cap gets applied by feel, trimming wherever is easiest, not where it matters least.

The hidden cost is the one you can't see: a "concise" answer looks clean and finished, so you never notice the qualification or the step that was quietly dropped.

The more honest move

You're conflating two separate things: getting the content right, and getting the length right. Separate them. Generate the substance first, then edit for length.

Content first, then a length edit [First, with no length constraint:] Explain this concept fully, including the parts that are easy to get wrong. [Then, once the substance is right:] That's the right content. Now cut it to about half, keeping the three key points and dropping the examples.

Length is best treated as an editing pass, not a generation constraint. When you cut after the fact, you can see what you're removing and decide whether it mattered. When you cap up front, the model decides for you, silently, and usually by feel.

Try this

When you want something short, resist setting the limit first. Ask for the full version, read it, then ask for a specific cut.

Name what to keep, not just how short to go: "keep the three main points, drop the background."

If you do need a hard length, treat the number as approximate and check it yourself. The model won't hit it exactly.

The principle underneath

The model approximates length and pays for brevity with substance you won't see it cut. Get the content right first, then edit for length.