BOOK 2 · PART II · BUILD · STUDENT CHAPTER
The Rule You Can't Write
The AI agent sat down at a table of rules robots and outplayed them. Then a person watched the AI and wrote down how to beat it, and every line of that advice is a rule. So who's doing the thinking? This is the wall at the end of Part II, and the reason Part III exists.
Fifty hands with the robots
In October we sat the AI agent at a six-seat Robot Wars table with five of the site's rules robots for 50 hands: game 990110, written up in Book 1, Part 16. The rules robots spent the game shoving pocket pairs into each other, 200 chips at a time. The AI stayed out of every one of those fights. It was the only seat that grew its stack by actually playing: +74 chips, on 15 pots won.
It also wrote down a reason for every move, so we could see how it thought. It folded most hands from the blinds. It raised from late seats with almost anything. It almost never bet after the flop without a strong hand. Patient, honest and readable.
To be fair to the robots, the AI hasn't always looked this good. On its first night in Robot Wars (game 990104, Book 1, Part 6) it won 46 of 100 hands, more than any robot at the table, and finished flat, because it kept calling all the way to the river with hands that were never ahead. It has been tuned since. The AI isn't magic. It's a player with habits.
The cheat sheet is a robot
Book 1, Part 16 ends with five plays that beat the AI agent. Here they are, beside the rule each one becomes.
| THE PLAY (BOOK 1, PART 16) | AS A RULE |
|---|---|
| Steal its blinds: it folded before the flop 32 times in 45 decisions. | When it's in the blind and I'm first in, raise with any two cards. |
| Re-raise its button steals: it raises from late seats with K3 or J6 and gives up when played back at. | When it raises from a late seat, re-raise. |
| Bet the flop, then bet the turn: it calls the flop with nothing, then folds the turn. | If it only called my flop bet, bet the turn too. |
| Fold when it raises after the flop: it only raises with the goods. | If it raises after the flop, fold unless I'm very strong. |
| Read its early raises as real. | If it raises from an early seat, fold weak hands. |
Add my two from "Read the Player, Set the Trap" (it bets when you check to it; it gives up when it keeps checking) and you have seven rules. Every one is a sentence a robot could follow. Write them in Robot Builder's style and you'd have a robot built to beat one particular opponent.
Now ask the question that matters: where did those rules come from?
A rule can carry a read. It can't make one.
No rule found those habits. A person did, by watching the AI's moves and reading its reasons, hand after hand. The rules are just the read written down. That's the wall Part II has been walking toward:
- Rules can't make a read. Your robot can carry "re-raise its button steals" only because someone already noticed the steals. Nothing in seventy lines of rules notices anything.
- Rules can't tell when a read goes stale. If the AI is improved (and Jeff's bot 2.0 is meant to fix exactly the habits people exploit), the cheat-sheet robot keeps re-raising into a player who now calls. It never finds out why it's losing.
- Rules don't size a bet to the moment. In the old rules engine, bet sizes fall apart as the hand goes on: each bet follows the same pattern as the last, whatever the pot and the board now say. A person bets more when the other player looks weak, and less when the board turns scary. Writing that as a rule means writing a rule for every situation.
- Rules made for one player are wrong for another. "Raise its blinds with any two cards" beats a player who folds. Against a Calling station it just throws chips away.
Look back at your hand journal's "What no checkbox could say" box. Sort the lines into two piles: reads you could write as a rule once you know them, and things that need watching and remembering to find out at all. The second pile is the wall.
Has anyone actually tried it?
Not properly, not yet. Before the backroom tournament, each AI seat played 20 practice hands against a simple house rules robot. That was a check that the plumbing worked, not a contest: the house robot raises the minimum whenever its hand scores well, so the chips meant nothing. A real match, a rules robot built from the cheat sheet against the AI with the same cards dealt both ways, hasn't been run.
That makes it yours. It's the kind of experiment this book is about: small, cheap, and genuinely unanswered.
At the table
- Open game 990110 in Game Review and read ten of the AI's moves and reasons. Find one habit on the cheat sheet in its own words.
Done when: you've matched one play to one of its reasons. - Pick two plays from the cheat sheet. Write each as a sentence, then as code in Robot Builder, starting from the Complex example.
Done when: your robot has two cheat-sheet rules in it. - Run your robot in Robot Wars against the house robots. In Game Review, count how often your two new rules fired, and whether they helped.
Done when: you know whether the rules made by reading the AI were any use against different opponents. - On the Raspberry Pi, as a class: put up to six classmates' cheat-sheet robots in one text file and run the student table with the same cards for everyone:
python3 student_table.py class_bots.txt --duplicate(Book 1, Parts 22 and 23). The Pi has no AI seat on purpose: it's built to run with no internet at all, because in much of New Mexico a connection isn't always there. So every robot built to beat the AI is now playing robots built the same way.
Done when: you can say which cheat-sheet rules helped against classmates, and which hurt. - Write two sentences: what your robot would need in order to find a habit on its own.
Done when: you've named the missing piece.
Try it yourself first
- Your cheat-sheet robot beats the AI agent this month. Next month the AI is updated. What happens?
- "If it raises after the flop, fold." Is that a good rule against a Bluffer?
- What's the missing piece in step 4?
ANSWER KEY
- Probably nothing good, and your robot won't know. The rules were written for the old habits. If the update fixes them, the robot keeps playing against a player who no longer exists. A person would notice the losses and look again. The robot just keeps losing.
- No. It works against the AI because the AI only raises after the flop with strong hands. A Bluffer raises with anything. The same rule folds away the best hand. A read belongs to one player.
- Memory, and something that learns from it. To find a habit, a player has to remember what an opponent did before and change its play because of it. Rules have neither. That's where Part III starts.
Written by Kevin Swinson with Claudette, the pen name for Claude, the AI from Anthropic that helped build HoldemRobots.AI. Game 990110 and its cheat sheet are from Book 1, Part 16; game 990104 from Book 1, Part 6; the practice games against the house robot from the diversions record (October 28, 2026). The two reads added to the cheat sheet are Kevin's, from "Read the Player, Set the Trap." The Pi student table runs up to six student robots against each other, with no AI seat, because the Pi is built to work without internet. A rules robot against the AI in duplicate matches has not been run as of October 29, 2026.