I Was Stuck in PU, and I Liked It.

You've seen the post. Somebody used one prompt, now they ship ten times faster, and here's the screenshot to prove it.

I went the other way. I told my AI to do less.

It's called Productive Underspecification. PU. Yes, it sounds like poo. I know. Give me a minute.

The genre has a tell

Every one of those posts sells the same thing, and it isn't really a prompt. It's volume. Paste this, get more: a longer answer, and the confident tone that makes it feel like real work happened while you got coffee.

Underneath there's a quieter tell. The prompt is a coupon. You use it once, screenshot the win, and by the next session your AI has forgotten it existed, so you paste it again. The output looks like a lot and lasts about a day.

Here's my problem with all of it. My AI was never stingy. It was too generous. It agreed with me about things I had wrong, and it did so in paragraphs.

So I told it to do less

The instructions fit on the back of a receipt. Tell me what you actually checked, and what you're just repeating back to me. When you're not sure, say so in a sentence instead of hiding it in five. And get shorter as you get less certain, so I can feel the difference.

That's the whole thing. PU. And honestly the name fits, because most AI output is exactly that, confidently.

My AI got quieter. It started saying "I haven't checked this" in the middle of answers that used to sail right past me. Once or twice it told me my idea didn't hold up. The first day it felt like a downgrade. By the end of the week it felt like sitting next to someone who knew the difference between what they knew and what they were guessing.

The gap is the point

Notice what the wall of text is for. It's covering a gap.

You asked something you don't fully know the answer to. That uncertainty is a small open space, an I-don't-know-yet, and it's uncomfortable to sit in. AI closes it instantly, in clean prose that's sure of itself, and the discomfort is gone. So is the one signal that was worth keeping.

That gap is where you do your actual work. It's the half-second where something feels off. It's the call only you can make, because you're the only one who knows what this is really for. When the AI fills it for you, it doesn't just answer the question. It quietly takes the question away.

So I stopped closing the gap. I left it open and looked at it. I got stuck in there, in PU, and the surprise was that I didn't mind. After a while I started liking it. The stuck part is where the thinking is. (If you want the long version of why agreeable AI costs you that judgment, we wrote it up in The Most Expensive Word in AI Is Yes.)

So I didn't want to lose it

Here's the catch with receipt-back rules. Paste them into a session and they die when the session does. The next morning your AI is blank again, and you're either re-typing the rules or quietly giving up on them.

That's the part we built fai for. Your AI forgets between sessions; fai doesn't. So the posture isn't something you paste back every morning. It's there when you sit down, and it keeps building on what you decided instead of resetting to whatever the model would have agreed to.

The Invitation

You don't have to take any of this secondhand. You can be stuck in PU yourself in about two minutes.

Install fai:

deno run -A jsr:@fathym/fai/install

Open your project and run it:

fai

Then open your AI coding agent - Claude Code, Cursor, and the like, not the terminal - and type /fai:meta:human. That's the posture: the AI does less, tells you what it actually knows, and hands the judgment back. Live in it for a day and see whether you like the gap as much as I did.

Build anything with AI. Keep everything. Evolve forever.

Start with fai - free ->