Karl's AI Watts publishes prompts and skill to remove AI-looking style from generated output

Karl's AI Watts (@aiwarts) posts six prompts for removing the AI-generated look from DeepSeek's new model and Codex output, covering redundant copy, boxed highlights, color, font sizes, one-message-per-screen layouts and animation purpose.

The author says the prompts target code-driven animations of the kind Claude Opus 5.5 produces. He also teases a video-editing Skill that he says has produced a ten-minute talking-head video and twenty case videos, with release planned for that week.

A follow-up report says he tested SenseNova-6.8-Flash on a 15-second code animation, a UI homepage and a Three.js 3D scene, then reran them with his vibecoding-humanizer skill. He says dropping unnecessary English captions and fixing color and type scales upfront reduced the generic look, and that adding Blender and explicit lighting fixed the 3D scene's floating objects and physics errors. The original post offered no results or comparison showing the prompts reduce AI-style output, and the follow-up's results are the author's own account.