Would I Have Typed This?
AI answers are verbose because the model needs its own words to keep working - not because the reader does.
There’s a question that works uncomfortably well: would I have typed this, if I had to write it to this person myself?
Not “is it correct.” Not “is it complete.” But: would I have bothered to write every single one of these sentences?
For most AI answers we forward, the answer is no. And not by a small margin.
The verbosity isn’t a bug
We’ve been complaining about the same thing for two years now: AI assumes things because we give it too little context. Asked too briefly, guessed wrong, answer unusable. Everyone knows that failure by now.
The same thing happens in the other direction. That AI answers run long is hardly news - everyone complains about it. But it gets filed as a matter of style. Almost nobody treats it as a context problem.
AI gives us a lot of context back. That’s not a quirk. The model needs its own words to keep working; the verbosity is functional. It’s just addressed to a machine.
And then we forward it to a person, unchanged.
Two boundaries, not one
Handing off to a model and handing off to a person are two different boundaries. We treat them as one.
For a person, the question is: what context does this one really need - and what can I take for granted? She sat in the same meeting. She knows the system. She doesn’t need the three opening paragraphs on fundamentals.
That’s not a question of form. It’s a question of substance. Not a prettier presentation - what’s left once you remove everything that isn’t needed.
Same fix, other direction
Some people say: then write it yourself. I’m less strict about that - doing it by hand isn’t the point. We already know how to fix the problem at the first boundary: give more context. The very same remedy works at the second one.
Tell the AI who the answer is for. What the recipient already knows, what you can take for granted, what he’s waiting for, what shape he needs it in - a Slack message isn’t a report. None of that lives in any model. Only you have it.
Then make it a standing criterion. The same way you write “no background explanations” or “ten lines maximum” into your prompt, this belongs in there: who it’s for. Once that’s a fixed part of how you prompt, there’s nothing left to trim by hand.
Which closes the loop: to have the AI fix the second failure, you have to avoid the first one. Too little context in and too much context out aren’t two problems. They’re one problem, twice.
The actual point
Research that used to take an hour now takes ten minutes. The gain is real. The only question is where it lands.
By default it lands entirely with the sender. He saves fifty minutes and forwards what he got. The recipient pays the difference - and that difference has been measured by now: people handed substanceless AI output spend close to two hours making it usable. He reads three screens instead of five sentences and digs out the part that concerns him.
Give some of it back. Not all of it. Twenty minutes usually does it.
Don’t optimize your time. Optimize your reader’s.