PM_tools — AI-Assisted Prediction Market Creator

A web application that uses multiple LLMs via the OpenRouter API to help users design well-defined prediction market questions through a multi-stage drafting, review, and refinement process.

How It Works

PM_tools guides users through a four-stage workflow to produce unambiguous, objective prediction markets:

Stage 1: Draft

The user provides a prediction market question, start/end dates, and selects a drafting model. The LLM generates a comprehensive market draft that includes:

• A refined, unambiguous question
• Detailed resolution criteria
• A complete set of mutually exclusive outcomes
• Edge case handling
• Potential sources for resolution

Stage 2: Multi-Model Review & Deliberation

Inspired by the llm-council "Structure D" deliberation model, multiple reviewer models critique the draft:

1. Independent Review — Up to 4 different models independently review the draft in parallel.
2. Deliberation — If 2+ reviews succeed, the first reviewer acts as "chairman" and synthesizes a consolidated deliberated review incorporating insights from all reviewers.
3. Human Feedback — The user can optionally add their own critiques before proceeding.

Stage 3: Update

The original drafting model incorporates the deliberated review and any human feedback to produce an improved draft that addresses the identified issues.

Stage 4: Finalize

The final draft is converted into structured JSON containing:

• Array of outcomes, each with resolution criteria
• Market start/end times in UTC
• Short description
• Full resolution rules
• Edge cases documentation

All sections are copyable to clipboard.

Architecture

src/
├── App.jsx                    # Main UI component and workflow orchestration (~600 lines)
├── main.jsx                   # React entry point
├── api/
│   └── openrouter.js          # OpenRouter API client with exponential backoff retries
├── hooks/
│   └── useMarketReducer.js    # Central state management via useReducer
├── components/
│   └── ModelSelect.jsx        # Reusable model selection dropdown
└── constants/
    ├── models.js              # LLM model definitions and defaults
    └── prompts.js             # System prompts and prompt builders for each stage

Key Design Decisions

• State management uses React's useReducer (via the useMarketReducer custom hook) rather than an external state library, keeping the dependency footprint minimal.
• Prompt engineering is centralized in src/constants/prompts.js with distinct system prompts for the drafter, reviewer, and finalizer roles, plus builder functions for each stage's user prompt.
• API resilience — The OpenRouter client (src/api/openrouter.js) implements automatic retries with exponential backoff (3 retries at 1s/2s/4s intervals).
• Model flexibility — 20+ models are available across 6 providers (OpenAI, Anthropic, Google, DeepSeek, Meta, Mistral). Default drafting model is GPT-5.1; default reviewer is Claude 3.5 Haiku.

Tech Stack

• React 19 with Vite for development and bundling
• OpenRouter API for LLM inference
• No additional runtime dependencies beyond React and React-DOM

Getting Started

Prerequisites

• Node.js 20+
• An OpenRouter API key

Setup

npm install

Create a .env file (or set the environment variable directly):

VITE_OPENROUTER_API_KEY=your_openrouter_api_key

Development

npm run dev

The app runs at http://localhost:5000.

Production Build

npm run build
npm run preview

Linting

npm run lint