1.7 KiB

Raw Blame History

Contributing to MiroFish BMNR

Thanks for your interest. Here's how to help.

High-Impact Contributions

These directly improve simulation accuracy:

Agent Personas

Edit src/data/agents.js. The persona field is a natural-language prompt sent to the LLM. The richer and more specific the persona, the better the agent reasons. Think of it like briefing a method actor — give them a backstory, investment philosophy, emotional triggers, and decision-making framework.

Stimuli

Edit src/data/stimuli.js. Add market events that affect BMNR. Each stimulus needs an id, name, cat (category), icon, impact (-1 to 1), and desc.

Prompt Engineering

Edit src/engine/prompts.js. This constructs the prompt sent to the LLM each round. Small changes here can dramatically change output quality. Test with multiple providers — what works for Claude may not work for GPT-4o.

LLM Providers

Edit src/providers/index.js. Add new providers by implementing the call(messages, options) interface. Return { text, raw }.

Development Setup

git clone https://github.com/mikema-rgb/BMNR-Mirofish.git
cd BMNR-Mirofish
cp .env.example .env
npm install
npm run dev

Pull Request Guidelines

Keep PRs focused — one feature or fix per PR
Test with at least 2 LLM providers if changing prompts
Update README if adding new features
Add your agent / stimulus to the appropriate data file, not to App.jsx

Code Style

No build tooling beyond Vite
Functional React with hooks, no class components
Inline styles using the T design token object (Wedge system)
Keep App.jsx as a working monolith — it needs to run as a Claude artifact

1.7 KiB Raw Blame History