I dump all my private notes into an LLM and tell it to build me a site

I’ve always found the concept of AI "hallucinations" fascinating. We constantly debate whether models are creative or just making stuff up, so I wanted to see what happens when you remove the guardrails entirely.

Tresbuchet is a publishing experiment. I drop a loose prompt into the agent along with a folder full of random text files, chat logs, and thoughts I’ve accumulated over the years. No specific instructions, no schema, no guidance on how to style it. I just tell it to "Go."

It reads the unstructured text and generates a completely new iteration of the site based on how it interprets the data.

It's been interesting tracking how different platforms (Claude, Codex, Gemini) handle the exact same mess of text differently. I've found that giving the models less instruction actually yields more interesting UI and copy results than trying to prompt-engineer a specific outcome.

I'm just building stuff to see what the models do when left to their own devices.

Curious what you guys think of the approach.