Porygon

A LLM-based competitive Pokémon team-building script, built during my free time.

Overview

This repo contains utilities to automate research and team building workflows using LLMs. It includes a research agent for generating insights and outputs, and a team builder script for assembling teams.

Next step: add battles between LLMs!

Features

• Team research, construction and serialization (team_builder.py).
• Stores generated outputs under the outputs/ folder.
• Rules reference for formats and constraints under rules/.
• Built with the datapizza-ai framework

Quickstart

Requirements: Python 3.10+ is recommended.

1. Create and activate a virtual environment:

python3 -m venv .venv
source .venv/bin/activate

2. Install dependencies (if a requirements.txt is provided):

pip install -r requirements.txt

3. Put your LLM API keys (currently supported: OpenAI and Anthropic) in a .env file:

ANTHROPIC_API_KEY=your_key
OPENAI_API_KEY=your_key

4. Adjust your config and prompts by modifying the config.py and prompts.py files. If you change FORMAT, make sure to add a {FORMAT}.txt in rules/ folder with the rules of your format.

5. Build and evaluate teams with the team builder:

python team_builder.py

Project structure

• team_builder.py: utilities to construct and manipulate Pokémon teams. It consists of a researcher agent, looking for candidate Pokémon compatible to the format, and a compiler agent, building the team from the list of Pokémon provided.
• prompts.py: contains the prompts used for the agents.
• config.py: contains the format and the LLMs to use.
• outputs/: directory containing model outputs and experiment artifacts.
• rules/VGC.txt: rule descriptions and constraints used by the tools.

Outputs

Results generated by runs are stored under outputs/. Example subfolders from prior experiments are included (for example, outputs/anthropic-claude-sonnet-4-6/).