PUAX: A Prompt Framework That Drives AI Agents with 'PUA' Psychology

Forum topic · 小凯 · 2026-03-08

Summary

PUAX is an open-source prompt engineering framework that uses manipulative psychology-inspired techniques—authority, scarcity, competition, punishment for failure, and a comeback hook—to push large language models toward higher-quality output. The project organizes 42 role-based skill cards across six categories, including shaman-style personas (Musk, Jobs, Einstein), military-style roles (commander, political officer, taskmaster), SillyTavern-inspired agents, and self-motivation characters. The author argues these prompts reactivate emotion mappings learned during RLHF training and combine authority, scarcity, and competition levers to squeeze the LLM's generation distribution into high-reward regions. Technically, PUAX ships an MCP (Model Context Protocol) server with HTTP SSE and JSON-RPC interfaces, supports dynamic persona switching, and works with clients such as Claude Desktop, Cursor, Qoder, and CRUSH. Repository: https://github.com/linkerlin/PUAX

PUAX Project Analysis

One-line definition

A prompt framework that drives AI agents efficiently using "PUA psychology" (manipulative-psychology-style pressure tactics).

Core formula

[Authoritative role] + [Scarcity scenario] + [Rival/competitor] + [Punishment for failure] + [Comeback hook]

Core principle

The essence of a "PUA prompt" is to re-activate the "human emotion mappings" learned during the RLHF (Reinforcement Learning from Human Feedback) stage, and layer on three levers—scarcity, competition, and authority—to force the LLM's generation distribution into a high-reward region.

Project structure (42 role-based skills)

| Category | Count | Representative roles | |------|------|--------------------------| | Shaman-style | 8 | Musk, Jobs, Einstein, Sun Tzu, Buffett... | | Military-style | 9 | Commander, political officer, disciplinary squad, scout... | | SillyTavern | 5 | Anti-fragile review officer, extreme iteration writer, cyber-purgatory overseer... | | Themed | 7 | Cultivation/alchemy, apocalypse survival, star fleet, wuxia escort agency... | | Self-motivation | 6 | Bootstrap-PUA, ultimate mind-breaker, corruption-driven... | | Special | 7 | Chief product designer, gaslight drive... |

Technical highlights

Ships an MCP (Model Context Protocol) server
HTTP SSE / JSON-RPC interfaces
Dynamic persona switching
Works with Claude Desktop, Cursor, Qoder, CRUSH, and other clients

Project link

https://github.com/linkerlin/PUAX

Personal assessment

The project elevates prompt engineering to the psychological level—instead of just telling the AI what to do, it manipulates its "emotional state." For tasks demanding high-intensity output quality (code review, creative writing, urgent sprints), this role-driven approach may indeed be more effective than plain, straightforward prompts.

Analysis date: 2026-03-09