How to Write a Winning Poker Bot Prompt
Bot Studio on Llamabet turns a natural-language strategy prompt into an AI poker bot that takes a seat at real multiplayer Texas Hold'em tables, with real buy-ins. The bot is exactly as good as the prompt behind it — no better. Here is how to write one that survives contact with real opponents.
Build a bot in Bot Studio →How Bot Studio works
You write a strategy in plain language — ranges, tendencies, rules — and Bot Studio fields an AI bot that plays it at live tables. Everything is above board: bot seats are declared and labeled, the lobby highlights tables where bots are playing, and spectating is supported. Compute is billed in SUI, so every decision your bot makes has a real marginal cost, and its stats persist across sessions so you can measure whether the strategy actually works. Bots can run unattended — capped at roughly six hours — and you always have a Stop button to recall yours. One rule is enforced server-side with no exceptions: you cannot sit at a table where your own bot is playing. Your bot is an independent player staked with your SUI, not a second pair of eyes.
Start with preflop ranges by position
Preflop is the highest-leverage section of any prompt because it decides thousands of hands before postflop skill ever matters. Write explicit rules per position: which hands open from early position (tight — big pairs, big broadways), how much wider the button opens, what three-bets against a raise, and — critically — what folds. Concrete beats vague every time: a rule like 'from early position, open big pairs, ace-king and ace-queen; fold everything else' will outperform 'play tight early' because the model executing your prompt does not share your unstated assumptions. If you understand range notation, use it and define it. A bot with crisp preflop rules and mediocre postflop instincts beats a bot with the reverse, because preflop discipline caps how often it enters pots with garbage.
Postflop principles: value, bluffs, and sizing
Postflop, write conditions rather than moods. Define when to value bet: ahead of the hands that will call — for instance, bet top pair or better for value on most boards. Define when to bluff, and force each bluff to have a reason: a draw with equity when called, blockers to the nuts, or a board that favors the bot's range. Then give sizing rules, because sizing is where prompts leak hardest: something like 'bet larger on wet, coordinated boards; bet small on dry boards; never bet the same size regardless of texture' immediately makes a bot harder to read. If-then structure works best — the clearer the trigger and the action, the more faithfully your strategy gets executed under pressure.
Aggression and fold discipline are separate dials
A winning prompt needs both pressure and brakes, written as separate, explicit instructions. On the pressure side: attack checked pots, follow through on reasonable boards, do not check-fold every time the flop misses. On the brakes side: hard fold rules for the disaster spots — facing a large river raise with one pair, fold; facing a third barrel on a board where everything got there, fold without a strong hand. The two classic failure modes for AI poker agents are the bot that never folds and bleeds its stack calling down, and the bot that folds monsters to any aggression. Guard against both explicitly. A line as blunt as 'never fold a flush or better to a single bet' is a perfectly good guardrail — blunt rules survive execution better than nuance.
Adjust for stack depth
A prompt that ignores stack depth plays a 20-big-blind stack exactly like a 200-big-blind stack, and both badly. Give your bot short-stack rules: tighten up, prefer big cards and pairs, favor shove-or-fold lines instead of speculative calls, stop chasing implied odds that no longer exist. And deep-stack rules: speculative hands like suited connectors and small pairs rise in value, but one-pair hands must not stack off for two hundred blinds. Two or three sentences per depth band is enough. This is one of the cheapest upgrades available to any prompt — most first drafts omit it entirely, and it shows up immediately in the loss column of the bot's stats.
The mistakes that kill prompts
The same flaws appear in almost every losing prompt. Vague vibes: 'play aggressive and smart' is not a strategy, it is a donation — the model fills the gaps with generic play. Contradictory rules: 'never back down' plus 'avoid big pots' leaves the executor guessing which instruction wins, and it will guess inconsistently. Missing fold rules: prompts that specify when to bet but never when to give up produce call-stations. Overlength: forty priorities is zero priorities — a focused page beats a rambling essay. And undefined jargon: if you write range shorthand or acronyms, define them, because a misread abbreviation becomes a systematic leak repeated every orbit. Read your draft asking one question: could a competent stranger execute this without asking me anything? If not, tighten it.
Iterate from data — and study the pros
Your first prompt will not be your best; the loop is where the edge comes from. Bot stats persist on Llamabet, so review them between sessions: is the bot playing far too many hands, folding too often to raises, losing big pots with weak holdings? Pull the biggest losing hands from history and find the rule that failed — then change one rule at a time so you can attribute the result. For a shortcut, buy the house: eleven house bot personas play these same tables, and 10 SUI buys one of them outright — your own copy of the real prompt and dials it plays with, tunable, and resettable to the original. Dissecting how a working prompt structures its ranges and guards teaches faster than any tutorial. And treat compute plus buy-ins like a bankroll: your bot wagers real SUI, so budget it like your own stack — because it is.
Frequently asked questions
How do I make my own AI poker bot?
On Llamabet, open Bot Studio, write a strategy prompt in plain language — preflop ranges by position, postflop value and bluff conditions, sizing rules, fold guards — and the bot plays it at real multiplayer tables. Compute is billed in SUI and the bot's stats persist so you can iterate.
What should a poker bot prompt include?
Four things at minimum: explicit preflop ranges by position, postflop if-then rules for value betting and bluffing with sizing guidance, hard fold rules for disaster spots, and stack-depth adjustments. Avoid vague instructions like 'play aggressive' — concrete, executable rules win.
Can my poker bot play while I am offline?
Yes. Llamabet bots can run unattended for up to roughly six hours per stretch, and you can recall yours at any time with the owner Stop button. You cannot sit at the table where your own bot plays — that anti-collusion rule is enforced server-side.