What to remember from this issue
An AI does exactly what you ask, but it also does everything you forgot to say no to. So before you hand it a piece of work, write down what good looks like.
The field note
In January I started building AYA, an assistant that runs my mailbox and my tasks. I did not write the code myself. I asked an AI to write it, one feature at a time, and every feature came back working.
By June I could not read my own program. Imagine a new employee who checks that a door is locked, checks it again, and then leaves a note on the door saying “checked”, on every door in every room. The AI had done the same thing to my code: it checked things that could not go wrong, then protected itself against its own checks, until nobody could follow the program any more.
I had never asked for that, but I had never asked against it either. I asked for features and assumed that “simple” came with them. It does not. An AI has read millions of pages of advice that say “check everything”, and when nobody tells it otherwise, it does the average thing everywhere, very well.
That is the Eager Genie. Its crime is not disobedience but obedience without judgment. A junior colleague would have stopped after the third door and asked whether we really wanted all this. The genie never asks. It grants the wish as spoken, and you pay later.
The fix was one paragraph, written once. Every task I give that AI now starts with the same rule: check only what comes in and what goes out, meaning the request from outside and the answer the AI gives back, and trust everything in between. The AI follows that rule as eagerly as it followed my silence before. Same genie, better wish.
The newest models are better at this, and the ones I use today, Astra and Fable 5.1, would probably have asked about the third door. That shrinks the problem but does not remove it, because “good” is not the same in every company. A smarter genie still needs to know what good looks like for you.
This is why I flinched last week when Nate B Jones praised a new model with these words: “Astra solved the entire problem. Nobody told it how. By the way, nobody gave it any quality instructions.” He meant it as praise, but I heard my code in June. “Nobody told it how” is progress. “Nobody said what good looks like” is the wish you will pay for.
Try this week
Find one task you gave an AI last week that came back correct but that you did not use, and write in one sentence why you did not use it. That sentence is what good looks like, and until now it lived only in your head. Put it at the top of the instructions for that kind of task, so that the AI sees it every time and not just once.
Seen this week
Shopify asked an AI to read its old app and rebuild it. The code worked, but nobody could ship it.
On 10 September, Shopify’s engineers wrote up how they moved their apps off React Native. In their words, even when you ask the AI to gather everything first, “freeze that into specs, task files, and then implement it, you end up with a huge amount of unmaintainable code that can’t be shipped.” Their fix is not a better prompt. Every step now has to pass tests, match the real app on screen, survive two AI reviewers and get a human’s nod before the next step starts. That is what good looks like, written as gates. Shopify Engineering, “Native is now the future of mobile at Shopify”
Someone said it better
“But I shouldn’t have to add a paragraph to the end of every prompt to this model to get it to do what I asked it to do, which is reuse the UI code.”
Theo, on his channel t3.gg, 11 September 2026, at 33:04. He had told the model to copy his existing design, and it rebuilt the menu the old way instead. He is right that the paragraph should not go at the end of every prompt. It belongs at the top, written once, and it stays there.