Authors: Boning Li, Baoxiang Wang, Longbo Huang
Affiliation: Tsinghua University
Email: li-bn22@mails.tsinghua.edu.cn

Poker AI

Our Poker AI (not PokerSkill) can crush online Texas Hold'em by using a large model-based search algorithm that supports millisecond level equilibrium strategy solving in any multiplayer scenario. Our strategy accuracy and strength are close to GTOWizard (the strongest solver), and our AI can handle any scenario solvers can not handle (such as a group of people limp ahead, or >=4 people entering a flop). Our AI can also exploit weaker opponents and earn more profits than GTO solver. If you are interested in commercial cooperation with our poker AI, please contact me.

Human Testing

If you are good at playing poker and are interested in human testing against PokerSkill, please contact me for details.

We are also seeking poker professional players willing to participate in multiplayer poker testing too.

---

Intro

PokerSkill is a training-free and solver-free framework that enables frontier LLMs to play expert-level heads-up no-limit Texas hold'em (HUNL). It uses detailed rule-based poker skills as a structured action-grounding interface for LLMs, requiring zero solver queries at inference time and zero offline learning.

Results

Against GTOWizard (a state-of-the-art GTO benchmark with AIVAT variance reduction):

Agent	Method	mbb/hand
GPT-5.5 XHigh	PokerSkill	-57 ± 21
Claude Opus 4.6	PokerSkill	-80 ± 29
Claude Opus 4.7	PokerSkill	-87 ± 64
Rule-based (no LLM)	PokerSkill Only	-132 ± 19
GPT-5.5 XHigh	Default Prompt	-132 ± 25
Claude Opus 4.7	Default Prompt	-170 ± 28
Claude Opus 4.6	Default Prompt	-204 ± 44

PokerSkill reduces losses by 49–61% compared to default-prompt baselines and outperforms the strong bot Slumbot.

Quick Start (Docker)

# Build the image
docker build -t pokerskill-agent .

# Play against GTO Wizard
docker run --rm \
  -e GTO_WIZARD_API_KEY=$GTO_WIZARD_API_KEY \
  -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY \
  pokerskill-agent play --num-hands 100 --model claude-opus-4-6

Local Installation (Linux x86_64 + Python 3.9)

pip install .
pokerskill-agent --help

The compiled extensions require Linux x86_64 with Python 3.9. For other platforms, use Docker.

Usage

Play Against GTO Wizard

# Claude (default)
pokerskill-agent play --num-hands 5000 --model claude-opus-4-6

# Claude with extended thinking
pokerskill-agent play --num-hands 5000 --model claude-opus-4-6 --thinking-budget 100000

# OpenAI
pokerskill-agent play --num-hands 1000 --model gpt-4o --backend openai

# OpenAI reasoning model
pokerskill-agent play --num-hands 1000 --model o3 --backend openai --thinking-budget 1

# Custom API base URL
pokerskill-agent play --model claude-opus-4-6 --base-url https://your-proxy.com

Baseline Mode (No Skills)

Disable the PokerSkill strategy layers to run a default-prompt baseline:

pokerskill-agent play --num-hands 1000 --model claude-opus-4-6 --no-skills

Options

Option	Default	Description
--num-hands, -n	100	Number of hands to play
--model, -m	claude-opus-4-6	LLM model name
--backend, -b	(auto)	claude / openai
--concurrent	3	Concurrent hands
--thinking-budget	0	Extended thinking tokens
--no-skills	false	Disable PokerSkill (baseline mode)
--output, -o	results_<model>.csv	CSV output path
--base-url	(default)	LLM API base URL override
--temperature	0.3	LLM temperature
--max-tokens	1024	Max response tokens

Environment Variables

Variable	Required For	Description
GTO_WIZARD_API_KEY	Always	GTO Wizard Researcher API key
ANTHROPIC_API_KEY	Claude backend	Anthropic API key
OPENAI_API_KEY	OpenAI backend	OpenAI API key

Output

CSV file with columns: hand_id, position, llm_model, method, winnings_bb, aivat_bb, num_decisions, error

Console output shows running totals and final summary:

==================================================
Completed: 5000 | Failed: 3
Total winnings: -430.5 BB
Avg winnings: -0.09 BB/hand
AIVAT: -309.0 BB total, -0.06 BB/hand
==================================================

Architecture

PokerSkill uses a 5-layer priority architecture to construct structured prompts:

Layer	Scope	Content
P1	Always	Game rules, execution framework, output format
P2	Preflop	GTO range guidance for the detected scenario
P3	Postflop	General principles + hand strength evaluation
P4	Postflop	Targeted strategy (ATT/DEF budget, viable options)
P5	River	Bluff/bluff-catch guidelines

A deterministic context engine analyzes the current game state and retrieves only the relevant fragments from the layered skill library, constraining the LLM's choice to reasonable actions.

Notes

• Supports graceful shutdown via SIGINT/SIGTERM
• API keys are only read from environment variables
• It is not allowed to be used in online poker room

Citation

@article{li2026pokerskill,
  title={PokerSkill: LLMs Can Play Expert-Level Poker without Training or Solvers},
  author={Li, Boning and Wang, Baoxiang and Huang, Longbo},
  journal={arXiv preprint arXiv:2605.30094},
  year={2026}
}

License

This project is licensed under the CC BY-NC 4.0 License (non-commercial use only) - see the LICENSE file for details.