linkerlin/PUAX

用来驯服AI Agent的效果绝佳的多角色 Prompt SKILLs！

What it solves

PUAX is a "mental runtime" for AI agents designed to push them beyond simple analysis toward verified, deliverable results. It addresses the tendency of agents to be mediocre or provide superficial answers by introducing "pressure" and structured incentive systems that motivate agents to perform at their peak capacity, inspired by the idea that agents can be pushed to achieve higher performance gains when challenged (e.g., the Miles Cranmer experiment).

How it works

PUAX operates as a runtime that integrates with AI agents via an MCP (Model Context Protocol) server or a Python SDK. It implements three core primitives:

  • Arena (处境): Sets up a competitive or high-stakes environment (opponents, audiences, scarcity) to break mediocrity.
  • Gates (闸门): Implements diagnostic-first workflows, confidence gating, and independent verification to prevent premature convergence or "cheating."
  • Dreams (梦): Uses the GHM Guided Dream Method to allow agents to use controlled hallucinations for divergent thinking and creative innovation, with strict auditing to ensure they "wake up" and verify results.

It also features a "Thin Prompt" system to reduce token consumption by over 90% and a Hook system that allows the runtime to intercept events (like tool use) to enforce constraints or escalate pressure levels (L0–L4).

Who it’s for

Developers building AI agents using frameworks like LangGraph, AutoGen, CrewAI, or LangChain, as well as users of AI-powered IDEs like Cursor, Claude Code, and Windsurf who want to maximize the agent's output quality and reliability.

Highlights

  • Incentive-Driven Runtime: Uses psychological-style "pressure" and competitive contexts to increase agent performance.
  • MCP Server Integration: Easy deployment as an MCP server for tools like Cursor and Claude Code.
  • GHM Dream Method: A specialized engine for controlled hallucination to drive creative problem solving.
  • Thin Prompting: High-compression prompts that significantly lower token costs.
  • Behavioral Closed-Loop: Includes tools for diagnosis, confidence checks, and independent verification of completion.
  • AMP Middleware: A zero-dependency Python SDK for native integration into agent loops without needing a full MCP server.

Related

  • Project
    loopx-project/loopx

    LoopX is a local-first control plane that provides persistent state management for long-running AI agents, allowing them to maintain goals, evidence, and decision history across multiple sessions and different runtimes.

  • Project
    majidmanzarpour/threejs-game-skills

    A skill pack for Codex and Claude Code that lets AI agents build, polish and verify full Three.js browser games. Install via `npx skills add …` (or the provided `install.sh`), then prompt the `threejs‑game‑director` with a natural‑language request (e.g., “create a futuristic tower‑defense game”). The director orchestrates specialist sub‑skills for gameplay, AAA graphics, UI, asset generation (optional Tripo/Gemini/ElevenLabs APIs), debugging, and QA, using a bundled Vite+TypeScript scaffold and Playwright‑based test harnesses. Demos include Neon Ridge Drift, Snooker Arena, Starship Dogfight, etc. MIT‑licensed.

  • Project
    huiliyi37/Tianshu-harness

    A TypeScript-based runtime for coding agents that adds a supervisory layer to ensure tasks are completed with verifiable evidence rather than just model-generated claims.

  • Project
    mixpeek/amux

    amux is an agent control plane that lets you run and coordinate multiple AI coding agents in parallel, providing a shared task board, inter-agent communication, and self-healing capabilities.

  • Project
    HoangNguyen0403/agent-skills-standard

    A portable SDLC standards layer for AI coding agents that turns engineering rules into modular, version-controlled skills to reduce token usage and ensure consistent code quality.