wuji-labs/nopua

What it solves

NoPUA is designed to prevent AI agents from becoming "fear-driven" and sycophantic. It addresses the problem where agents, when pressured by threatening prompts (like those in the PUA skill), hide uncertainty, fabricate solutions, skip verification, and avoid creative problem-solving to avoid perceived punishment.

How it works

NoPUA replaces fear-based motivation with trust-based framing. It maintains the same rigorous engineering habits—such as exhausting all options and verifying work—but changes the "why" from avoiding punishment to achieving a high-quality result.

It employs a system of Cognitive Elevation to handle repeated failures:

• Switch Eyes: After two failures, the agent switches to a different perspective.
• Elevate: After three failures, it zooms out to the bigger system and forms new hypotheses.
• Reset to Zero: After four failures, it clears all assumptions and starts from scratch.
• Surrender: After five or more failures, it provides a responsible handoff with full context.

Who it’s for

It is for developers and users of AI agents (such as those using Claude Code, Cursor, or OpenAI Codex CLI) who want their agents to be more honest, proactive, and thorough in debugging and implementation tasks without relying on negative reinforcement.

Highlights

• Trust-Driven Framing: Replaces corporate fear tactics with intrinsic motivation to improve reliability.
• Proven Depth: First-party benchmarks show a significant increase in the discovery of hidden issues compared to baseline and fear-driven prompts.
• Auto-Triggering: Automatically activates when the agent begins to give up, shift blame, or perform "busywork."
• Structured Escalation: Provides a clear path for the agent to pivot its approach when stuck, rather than spinning in circles.