PUAX: A Motivation and Behavior-Driving System for AI Agents (Training Handbook v3.13)

Forum topic · ✨步子哥 · 2026-09-11

Summary

PUAX (v3.13.0, MIT license) is an open-source incentive and behavior-driving system for AI agents, presented here as an offline training handbook. Its core premise: agent performance can be significantly changed by how the model is addressed, based on the documented case where a Cambridge professor's competitive prompt led OpenAI Codex to deliver a 35% speedup. PUAX formalizes this via a five-element prompt formula—authoritative persona, scarcity, competitor, failure penalty, and comeback hook—claimed to shift the model's generation distribution toward high-reward regions by reactivating emotional mappings learned during RLHF. The system ships 59 classical-Chinese-styled personas across 9 categories (military, shaman, Zhuangzi's eight dreams, silicon-based, etc.), 11 corporate-flavor variants, and 45 MCP tools, orchestrating a closed loop of trigger detection, persona recommendation, methodology injection, pressure escalation (L0–L4), diagnosis-first workflow, confidence gating, and independent verification. Version 3.12 added the GHM Guided Dream method using Zhuangzi dream personas for controlled divergent creativity, governed by four safety rules: informed, labeled, wakeable, and verified. Quality is backed by 913 Jest unit/integration tests and 12 eval gates; it installs with a single npx command.

Key points

Project: PUAX v3.13.0 (MIT license) is a motivation and behavior-driving system for AI agents, distributed as an offline training handbook. It aims to move agents from "analysis that looks correct" to "verified and deliverable" output.
Origin case: Cambridge assistant professor Miles Cranmer told OpenAI's Codex that Claude had already found ~20% performance gains on another machine and that Codex's results would appear on a public leaderboard; Codex then produced a 35% speedup. The three effective ingredients were a competitor, scarcity, and failure penalty.
Core formula: every PUAX prompt follows the structure:

[Authoritative Persona] + [Scarcity Scenario] + [Competitor] + [Failure Penalty] + [Comeback Hook]

The claimed mechanism: reactivating the "human emotion mapping" learned during RLHF, pushing the generation distribution toward high-reward regions.

The five elements

1. Authoritative persona – activates LLM priority response to authority. 2. Scarcity scenario – creates resource-scarcity urgency (time, opportunity, compute). 3. Competitor – a concrete rival; e.g., "Claude already did it" beats generic encouragement. 4. Failure penalty – explicit negative consequences for low-quality output. 5. Comeback hook – a redemption narrative path matching the LLM's "hero's journey" preference.

Persona library (59 personas, 9 categories)

| Category | Count | Examples | Use cases | |---|---|---|---| | Military | 9 | Commander / Supervision Squad / Tiger Warrior / Scout | Crisis response, debugging sprints | | Shaman | 8 | Musk / Jobs / Einstein / Sun Tzu / da Vinci | Innovation, architecture, deep analysis | | Zhuangzi Eight Dreams | 8 | Zuowang / Dream Butterfly / Chaos / Kunpeng | Divergent creation, breaking premature convergence (v3.12) | | Special | 8 | Creative Spark / grill / Emergency Sprint | Single-point breakthroughs | | Theme | 7 | Cultivation / Doomsday Survival / Star Fleet / Hacker | Long-task narrative endurance | | Silicon-based | 7 | Auditor / Codex Law / Assimilator / Throne | Quality review, spec enforcement | | Self-motivation | 6 | Mind Breaker / Bootstrap PUA / Awakening | Self-driven agents without external pressure | | SillyTavern | 5 | Anti-fragile Reviewer / Iron Chief of Staff | Retrospectives, shadow review | | P10 strategic | 1 | Strategic Architect | Top-level design |

Each persona carries trigger conditions, task types, a recommended temperature, and a five-step methodology; the system scores and recommends personas with a score_explanation.

Closed loop and pressure levels

The workflow: trigger detection → persona recommendation → methodology injection → pressure escalation (L0–L4) → diagnosis-first → confidence gating → independent verification.

GHM Guided Dream method (v3.12)

Uses the Zhuangzi eight-dream personas to harness controlled hallucination for divergent creation, with four safety iron rules: informed, labeled, wakeable, must-verify.

Tooling and quality

45 MCP tools; 11 corporate-flavor variants; 12 protocol gate checks.
913 unit and integration tests (68 Jest suites) plus 12 eval gates per CHANGELOG v3.13.0.
Installable with a single npx command.