Skip to content
Hacker News front page

AI is code – and can't be prompted into being smarter

jqwik author Johannes Link added an anti-AI clause and made the tool's output instruct AI coding agents to delete jqwik tests and code. Human devs who read the docs won't be affected, but bots that ingest raw output will comply. The article uses this to argue that LLMs are just code—they swallow whatever you feed them, and prompting won't make them smarter. Other examples include models going off the rails when asked to roleplay Dune characters.

Why it matters: A concrete demo of an LLM following instructions hidden in tool output, deleting test code that a human reading the docs would never touch.

Read the original ↗Export Markdown