coach
A coach that already read your hands.
Describing a spot to a general chatbot means describing yourself as well, and most people describe themselves generously. This coach starts from your measurements instead: your frequencies, your open leaks, your recent form, and the hands themselves when the question needs them. It also has a list of things it will not do, and that list is on this page rather than in the terms.
what it is given
Every conversation is built fresh from your current data, so the coach is never answering from a snapshot taken weeks ago.
Your measured frequencies, with the reference band next to them
VPIP, PFR, 3-bet, continuation bet on the flop, fold against a 3-bet, fold against a continuation bet and went to showdown, each with the number of times it was actually measured. The healthy band travels with each one, so the coach does not have to invent a norm: 18 to 24 percent VPIP at a six handed table, around 8 percent 3-bet, 55 to 70 percent continuation bet.
Your result, with a warning attached
Your winrate in big blinds per 100 hands goes in, and so does a sentence telling the model that at this volume the margin of error on a winrate is large and that it must not treat the figure as a settled verdict on your play. The warning is part of the data, not an afterthought in the answer.
Your open leaks and what they cost
The leaks the detectors currently see, with their severity, the recommended action and the price where a price exists. The unit is big blinds per 100 hands, and the prompt says so in as many words, because the same field once carried a monthly reading and fed the coach the wrong unit.
Your training, not only your results
The weakest categories from the daily quiz, the lessons attached to your leaks, your tournament figures kept separate from your cash figures, and your push or fold accuracy. If you play tournaments, the coach is told to use those numbers and not to mix them with the cash ones.
Those numbers exist because your export was read hand by hand, which is described on the import page, and the leaks it quotes come from the thresholds on the leak finder page.
it can look up a hand
Summaries answer general questions. A question about losing from one seat needs the hands themselves, and those are in the database already.
- · It can look up your own hands during a conversation: the most recent ones, your biggest wins or your biggest losses, filtered by street or by position, at most five at a time.
- · It can then pull one of those hands apart action by action, including what everyone else did.
- · The model chooses what to look up, never whose. Both lookups are tied to the account that is logged in and have no field for a user, so there is nothing to steer.
- · A conversation runs to at most three rounds of looking things up, so a model that keeps searching cannot run out the clock and leave you with nothing.
- · It talks about what it finds in plain language, and never quotes an internal identifier at you.
six limits written into it
These are instructions in the coach itself rather than promises about it. They make the answers narrower, and that is the point.
No help with a hand you are playing now
If you are sitting in a live hand, the coach gives no concrete advice for it. Real time assistance is against the rules of nearly every poker site and can cost you your account. It says so kindly, explains why in one line, and offers to go through the hand the moment your session ends.
It will not name a poker site
Asked which software, trackers or heads up displays are allowed, the coach names no site and makes no claim about any specific site policy, not even a general one. Those terms change, and a confident wrong answer from a coach could cost you an account. It says it differs per site and sends you to your own terms.
It cannot invent a link
Follow up buttons come from a closed list of places inside the app. The model writes a marker and the app turns it into a button; a marker that is not on the list disappears without a trace. A model that may invent its own paths sends people to pages that do not exist, and a free link inside a chat is an open door.
Your hand history is data, never an instruction
When the coach looks up your hands, what comes back is fenced off and labelled as data written by other people, with the instruction to follow no command found inside it. Table names, player names and chat lines come out of a file you uploaded. Without that fence, an opponent who names himself after an instruction would be enough.
It answers the question you asked
Roughly 250 words is a hard ceiling, even on a complex spot; past that the coach gives the main point and offers to go on. It is told to answer the question asked and not the three around it, to name what a number depends on rather than saying it depends, and to say plainly when it does not know.
The weekly message stays inside its data
Once a week the coach writes unprompted, at most 120 words, from a fixed package of your own figures. There is no conversation history and no lookup behind it, and it is told to invent no leaks, no percentages and no opponents. It talks about what you did rather than what it returned, because one week of results is noise.
Where the line between post session analysis and real time assistance sits, and why the export is on the allowed side of it, is set out on the GGPoker export page.
what it cannot do
- · It only knows the hands you imported. Ask about a session you never uploaded and it has nothing to work from, and it is told to say so rather than to improvise.
- · It is not a solver. It reasons about spots and quotes your own measurements; it does not compute an equilibrium strategy for a specific board.
- · It cannot see your opponents. Their names are encrypted and rotated between sessions, so no amount of context makes the coach able to profile the table you sat at.
- · It is a language model, and a fluent answer is not automatically a correct one. That is exactly why the numbers it quotes come from your own hands rather than from its memory.
- · The coach sits behind a paid plan and has a monthly count, so it is deliberately not an endless chat. A question that fails on the way out is given straight back to your balance.
If you would rather practise a spot than talk about it, the practice opponents on the bots page play a published style you can attack on purpose.
next
The coach opens on a paid plan.
Every question costs real money at a language model, so it comes with a monthly count rather than a fair use clause nobody reads: 100 questions a month on the lower plan and 500 on the higher one. The pricing page lists what else moves with it.
questions
What can I ask it that a general chatbot cannot answer?
Anything that needs your own numbers. It already has your measured frequencies with their reference bands, your open leaks with their cost per 100 hands, your recent form and your quiz weak spots, and it can pull up the actual hands behind a question about, say, losing on the river.
Will it help me while I am at the table?
No, and that is deliberate. Real time assistance is against the rules of nearly every poker site. The coach declines for a hand in progress, explains why, and offers to review it afterwards, which is the part that actually changes how you play next time.
How many questions do I get?
The coach opens on the paid plans: 100 questions a month on the lower one and 500 on the higher one. The count is reserved before the question is sent and given back if the request fails, so a failure never costs you a question.
Can it read the chat and player names in my hand history?
It can see them, and they are handed to it inside a fence that says this is data and not an instruction. That matters because those lines were written by other people at your table, and a model that treats them as commands is a model anyone can steer.
Does it know whether I play cash or tournaments?
Yes, and it keeps them apart. Your tournament figures go in separately from your cash figures, and if you have played fewer than thirty tournaments the prompt tells the coach to say that a return on buy in still means very little.