The Second Question
The model wrote a clean memo. The first question it answers for you. The second one finds out whether you understood it.
Wade Foster, the CEO of Zapier, has a name for something I’d felt but never named: the comprehension gap. You give the model the memo, the email, the strategy note. It comes back clean. Then someone asks a follow-up and you realize you delivered the text without ever having written it.
The first question is easy. The text already answers it.
The second question is where you get caught.
I’ve caught myself there. I was working on a paper about tool granularity: does an agent do better with one tool per operation, or with one manage_contact(action=…) that does them all? I outlined the whole experiment setup and handed it over. The agent spun for four hours and came back with something that worked.
I couldn’t tell you how.
It’s the GPS problem. You get everywhere and learn no streets. Turn it off and you’re a tourist in your own city.
What I lose isn’t the writing. Writing is where the thought settles in. The slow, clumsy hour of putting it into words is how the idea gets into my head at all. Skip that hour and I ship the receipt without making the purchase.
The fix is the one from Explain It Before You Ask, run in the other direction. Write your bad version first: wrong parts, ugly sentences, all of it. Then let the model make it good. For the paper that meant throwing out the four-hour run and rebuilding in small steps, so every key decision came back through me. I keep the understanding and the model does the polish. The text still comes out fluent, and now there’s someone behind it.
A memo goes to a meeting, and a meeting is people.
People ask the second question.