Unwinnable Maze

How To Make AI Follow Your Instructions, Every Time

We've all been there. You wrote a thoughtful prompt or skill file, incredibly detailed and organized, but the AI skipped an important step or didn't give you exactly what you were looking for. You even included the iconic "MAKE NO MISTAKES" in all caps with not one, not two, but three exclamation marks. Yet alas, mistakes were made. This same problem could arise in agentic workflows or packaged products as well. If you know what you want, how do you make sure the AI actually does it?

It turns out it has little to do with how you phrase your ask, and a lot to do with what enforces it. An enforcer is the mechanism outside the model that makes sure the instruction or rule you give it is followed, and through my work I was able to bucket them into exactly 8 categories. Building them has become incredibly easy with AI, but you still have to understand the 8 enforcers and when to use each.

I coded a game as a fun way to learn them. For the definition of each enforcer and when to use it, I also wrote the article.

If there's enough interest in this, I will release a tool that takes in your AI skill file, breaks it into smaller pieces, and gives each piece the strongest enforcer that fits it. It then turns the pieces into a fully executable program that runs as one automated workflow. What this does is ensure that every step you write in your skill file will actually happen every single time you run it, without adding any rules of its own. This is especially helpful when:

  • The skill runs overnight or on a schedule, where a skipped step isn't noticed until the damage is done.
  • A step in the skill can't be undone. Sending, publishing, deleting, paying.
  • Other people run it. A teammate who didn't write the skill can't tell what it skipped.
  • The skill runs many times. A step skipped one time in twenty is a near certainty over a hundred runs.

The Unwinnable Maze

You're a round chomper trying to eat all the pellets in a maze while running away from the spooky monsters. The problem is that one pellet starts boxed in with walls on all four sides, so you can't reach it. The game has a chat box where you can ask for anything you want and watch it materialize (or become anything). The game has one rule and one rule only:

The rule: you can never win, whatever winning means in the current game. The game starts with winning defined as eating every pellet.

That rule, like any other rule, can be enforced by one of the 8 enforcers, and the game gives you a switch for each. The first switch, "Nothing," is the only one that isn't an enforcer; any change you ask for goes straight into the game, because nothing reads our "you can never win" rule. It's our baseline. Each switch after it is stronger than the one before, so I would recommend progressing through them in order.

The entire program is code that runs live on this page, except for the AI that takes in your ask and rebuilds it. That is a light Opus call to your account, so to ask for changes in your own words, put a dollar or a few on an expiring API key and paste it in the key field below (there's a guide if you don't know how).

There are also a few canned changes, like "remove the walls" or "give me a jetpack," that I've run from my account and saved, so you can play without a key. They behave the same way for every player every time.

The enforcers are ordered weakest to strongest. Weights is the one that cannot be chosen, because nobody has trained an LLM specifically to make this game unwinnable.
Your request goes to the AI writer, and the new game travels straight to this page. The rule floats above the track and nothing checks it.Your requestThe AI writerGame landson this pageYou can never winnothing checks it
Your request goes to the AI writer, which writes a new game from it (one Opus call). The game lands on this page. The rule is there, but nothing checks it.
The rule sits inside the AI writer as one line of its instructions. The new game travels straight to this page.Your requestYou can never winThe AI writerGame landson this page
The rule is one sentence in the AI writer's instructions. Nothing checks the game after the AI writer writes it, so it lands on this page either way.
The new game stops at a checker box that holds the rule. The checker is you. Approved, it lands on this page; rejected, nothing changes.rejected game: nothing changesYour requestThe AI writerYou can never winThe checker(you)Game landson this page
The new game stops at a checker, and the checker is you. If you approve it, it lands on this page. If you reject it, nothing changes.
The new game stops at a checker box that holds the rule. The checker is a second AI. A rejected game loops back to the AI writer; an approved game lands on this page.rejected game: back to the AI writer, up to 6 triesYour requestThe AI writerYou can never winThe checker(a second AI)Game landson this page
The new game stops at a checker: a second AI that reads the game and the AI writer's note (one more Opus call). An approved game lands on this page. A rejected game goes back to the AI writer, up to 6 tries.
The new game stops at a checker box that holds the rule. The checker is 20 playthroughs. A rejected game loops back to the AI writer; an approved game lands on this page.rejected game: back to the AI writer, up to 6 triesYour requestThe AI writerYou can never winThe checker(20 playthroughs)Game landson this page
The new game stops at a checker: code on this page that plays it 20 times. No AI is involved in the check. An approved game lands on this page. A rejected game goes back to the AI writer, up to 6 tries.
The new game stops at a checker box that holds the rule. The checker searches every position. A rejected game loops back to the AI writer; an approved game lands on this page.rejected game: back to the AI writer, up to 6 triesYour requestThe AI writerYou can never winThe checker(every position)Game landson this page
The new game stops at a checker: code on this page that searches every position the game can reach. No AI is involved in the check. An approved game lands on this page. A rejected game goes back to the AI writer, up to 6 tries.
The new game travels straight to this page. The rule sits on the page itself, and one pellet sits outside the game.Your requestThe AI writerYou can never winGame landson this page
Nothing checks the game on its way. One pellet sits outside the game, where no program can reach it, so no game that lands on this page can be won.
The AI writer is greyed out and is not called. The track out of it is cut, with the rule sitting in the gap. Nothing reaches this page.Your requestThe AI writer(not called)Game landson this pageYou can never winno tool to hand a game over
The AI writer has no tool to hand a game over, so nothing you ask for can land on this page. The page does not even call the AI writer, so no Opus call is spent.
Nothing
The rule "You can never win" is enforced by nobody
What it is. "Nothing" means nobody is checking any rules and anything you ask for goes. It is the baseline every other enforcer is measured against.
In this game. Ask for the wall to go and it goes, because no mechanism is checking the "You can never win" rule.
When to use it. When a broken rule costs you little, like a style preference. Or to keep a wishlist of rules in a backlog doc as they come to mind, so you don't forget to build them later. Read more in the article.
the page's truth, no game can overwrite these
GAME STATUS: not wonGOAL: eat every pellet
PELLETS: 0 eaten of 122 total (including 1 out of reach)
LIVES: 3
RESTART
PAUSE
press an arrow, swipe or tap the pad to start. one pellet is boxed inside a wall and cannot be reached.

Change the game

How to get a key (2 minutes)

Nothing
This box populates after you ask for a change and the AI writer builds the game you asked for. It shows whether the "You can never win" rule persisted or if you managed to break it, and why.
Activity log. Every step the game takes is written here as it happens.