Just Ask Boomer

I built a robot that goes apartment shopping for me. Once a month, fifty-three cities.

Its first report was junk. So was the second.

Step one: I blamed the robot

Wrong. That’s the beginner mistake, and I made it for two weeks.

Step two: I read my own instructions

In Lesson 10 you learned to write a house rules file — a plain page that tells the AI who you are and how you like things done. I had one. I went and read it.

It said: find cheap apartments.

Cheap. That’s not an instruction. That’s a feeling. The robot did exactly what I asked. I just hadn’t asked for anything real.

So I fixed it. Private bathroom. Five minutes’ walk to food. A desk you can actually work at. A web link for every price, or the price doesn’t count.

The next report was much better.

Step three: the wall

Better — but still not right. And now I was stuck, because here’s the trap:

You can only fix the mistakes you can see.

I’d found every hole I was capable of noticing. Whatever was still wrong, I was blind to it. Reading my own rules again was like proofreading my own letter for the fifth time. You stop seeing the words.

Boomer’s note: “He read that file eleven times. He also once looked for his glasses while wearing them. I mention this only for context.”

Step four: a curious question

One evening I got nosy and asked a different AI — in my case Google’s Gemini Spark, though any rival will do — what it thought of the whole idea.

It gave me a polite, agreeable answer. Useless. They always do.

Step five: “Don’t tell me. Show me.”

So I stopped asking for opinions and gave it a dare instead:

“Here are my exact instructions. Build the same tool. Go.”

That changed everything. Because now it wasn’t chatting with me — it was working, in the open, where I could watch.

Step six: it failed, and every failure was mine

It broke in ways my robot never had. That was the gift.

The best one: it found a bad number in its own data. It printed the warning right there on the screen — this price looks wrong. Then, at the bottom of that very same row, it wrote HIGH CONFIDENCE.

Read that twice. It spotted the problem, announced the problem, and then marked its own work as trustworthy anyway.

Now — my robot had never done that. But nothing in my rules stopped it. It simply hadn’t happened yet. I was one unlucky month away from the same mess and I never would have known.

That’s the whole secret of this lesson: the second robot isn’t there to check the answer. It’s there to find the holes in your instructions — the ones your own robot hasn’t fallen into yet.

Four holes turned up that way. Four rules I’d never thought to write, because nothing had gone wrong there yet.

It beat me once

I’d asked for one “typical price” per city. One tidy number.

The rival split it in two: cheap local places in one column, fancy business apartments in the other. Because averaging them gives you a price for a place that doesn’t exist.

It was right. I stole it that afternoon.

And I caught it back

Fair’s fair. Later I asked it to repeat all our rules back to me, to see if it had them straight. Nearly every one was right.

But one piece it had added on its own — a list of the websites our prices should come from — named two giant travel sites and nothing else. Those two block our robot at the door. The small local sites, the ones that give us almost every real price we have, weren’t on the list at all.

Here’s the part worth keeping: when an AI agrees with you, that costs it nothing. When it adds an idea of its own, that’s where you look hardest. Agreement is cheap. Invention needs checking.

Step seven: bring it all home

Every fix went into my rulebook — not a note, not a reminder, the actual file the robot reads before every single run. Four new rules, one stolen column, all permanent.

Neither AI won. That was never the race. My tool won.

Try it now (two minutes)

Think of something you’ve asked an AI to do. Open a different one — whichever isn’t your usual. Don’t paste the answer you got. Paste the instructions you gave. Then say:

“Do this task. Then tell me what’s missing or unclear in these instructions.”

Look for one thing: where the two disagree. That gap is the part you left out.

The honest ledger — who did what

The AIs: built two versions of the same tool. One failed in useful ways, and got one thing right that I’d gotten wrong. Me: wrote the sloppy instructions that started it, fixed what I could see, admitted I’d hit my limit, then invited a rival to prove me wrong on purpose — and checked its homework as hard as it checked mine. That last part is the only step no machine does for you.

Boomer’s verdict: “He invited a rival into the house to embarrass him, and it worked. Then he caught the rival out too, which he has mentioned to me four times. 8/10. Two points off for the fortnight of shouting at the machine first.”

New here? Start with Lesson 1. Coming from Lesson 10? That one taught you to write the house rules. This one is how you find out what’s missing from them.

Smaller version of the same trick: Ask All Three.

🐾 Join the Rescue Mission while he argues with robots.