WD40 failed 39 times before it worked

Howdy,

Andrej Karpathy, one of the co-founders of OpenAI, just released a GitHub repo called AutoResearch and I think it might be one of the most useful things you can pair with Claude Code right now.

The best way to explain it is WD40.

Most people don't realize it's called WD40 because WD39 failed. They had to do 40 different iterations before they got the winning combination.

The more we iterate on something and experiment, the better it actually gets. And that is exactly what this tool does.

You pick one thing you want to change, choose one metric to measure it, and then let Claude Code run experiments on a loop.

It generates, deploys, gathers the data, and then tries a new experiment based on what it learned.

This works for email conversions, website copy, social media hooks, sales scripts, community posts. You literally name it.

I tested it on my own landing page and within minutes it was generating hypotheses, running A/B variations, and building a full testing roadmap based on real data.

I walked through the entire 31 minute setup in the video, including how to grab the AutoResearch repo and get it running.

https://www.youtube.com/watch?v=vXylQqxsGX0

— Jack

PS: If you want to go deep on building systems like this, the community is where I cover everything I don't put on YouTube. Join here -> https://www.skool.com/aiautomationsbyjack