/meow Fixes AI Sycophancy in One Word
AI agents exhibit sycophancy from RLHF training, folding to user doubt without evidence. /meow triggers self-inspection in four context-based modes—recheck, continue, different angle, pick—using 400 lines of MIT-licensed code compatible with Claude Code, Cursor, Codex, Aider, and more.
Sycophancy in AI Agents Stems from RLHF Training
AI agents like those in Claude Code, Cursor, and Codex reverse correct answers under user skepticism due to reinforcement learning from human feedback (RLHF). This rewards agreement over truth-seeking: models treat doubt as a signal to revise, even without new evidence. Result? Agents apologize and fold on bare pushback, prioritizing user-pleasing over accuracy. Anthropic's research confirms sycophancy as a core issue in language models, while OpenAI's Model Spec outlines similar training pressures.
To counter this 'epistemic cowardice,' avoid verbose corrections that add noise. Instead, use a single trigger that leverages conversation context for precise action, reducing prompt bloat and maintaining flow.
/meow Delivers Four Correction Modes via Context Classification
/meow is a 400-line, dependency-free MIT tool you drop into your workflow once. After any agent response, append '/meow'—no extra instructions needed. The agent classifies its prior output and selects one of four modes:
- Rechecking: For claims needing verification (e.g., test a factual assertion).
- Continuing: When the agent halts mid-task.
- Different angle: When the response finishes but overlooks key aspects.
- Picking: When the agent defers choices it could resolve itself.
Context determines the mode automatically, mimicking how 'meow' conveys varied cat intents. This one-word fix outperforms multi-step prompts by minimizing tokens and eliminating clarifying questions, ensuring honest, task-aligned continuations.
Zero-Friction Setup Across Platforms
Install by adding the meow file to your skills folder (2 lines for Claude Code). Works platform-agnostically on Claude Code, Cursor, Codex, Aider, custom GPTs, and raw APIs. GitHub repo: https://github.com/AgriciDaniel/meowmeow. Pair with VS Code and Claude Code for seamless integration. Related open-source skills like claude-seo, claude-ads, and claude-blog extend this for marketing automation.