I built an agent to analyze my past poker hands. It showed me how to stop losing.

I play low stakes poker online. I am average at best. I wanted to up my poker game so that I could change my strategy to start trying to be profitable.

I wanted to read all my hands and build an agent to see and analyze the way I played to lose money. The analysis would run in real deterministic code, so the numbers came out exact and repeatable.

I started off by looking at my poker website's hands played.

I was hoping to find the history of my hands completely in a downloadable file. What I found instead was a hard-to-download hand-by-hand analysis per day.

I wanted to first prove out that there was enough information to recreate the hands that I played. I took the hand info from a couple of hands I had just played and put them into Claude Code.

It was able to accurately tell me the hands that I had.

It reasoned through the hands and told mine apart from everyone else's, representing my play accurately.

Now it was time to see if I could find a way to download my entire playing history.

I sniffed out the backend request behind my login and downloaded my whole hand history.

I pulled 1,285 hands over the last 60 days. I was down $8.06.

One clear area of my game needed work.

I was limping pre flop (calling before seeing cards).

I played 819 hands where I limped pre flop. 6 out of 10 of these hands a normal player would just fold. Most of the damage came from calling with them instead of folding:

The other 130 I raised instead of folding. Those actually won $6.76, but that's luck over a small run.

I found out I was losing a lot of money to the rake (the money the casino takes out of your winnings).

$20.66 was raked out of the pots I won. Without that rake, I’d have been up $12.60 for the session instead of down $8.06.

It was crazy to see how much that rake was taking from my profits.

So my edge is already positive. The rake is what's putting me underwater. That matters, because "play better" and "play somewhere the rake is lower" are two different fixes, and I'd have wasted a month on the wrong one.

The first report said I saw a flop in 75.6% of hands, while also saying I only played 64%. You can't see more flops than hands you played.

If you plan on doing something like this yourself, you need to triple check the outputs you are getting. Do not just take the outputs of the agent at face value. Validate with test cases and examples.

After this analysis I was able to play some more hands over the last few days and I had my agent analyze my play. This time I was up $6.21, and the limping leak was gone.

From here, the natural next build is an AI trainer that drills the exact spots you keep misplaying.

If you play poker (or any repetitive game) and want to run this kind of analysis on your own hands, give it a shot. I’d love to hear what you find, drop me a note at [email protected].