Delete your CLAUDE.md (and your AGENT.md too)

I knew being lazy to type up a CLAUDE md would pay off

I only write stuff there after I see the agent fail on something to give minimal context not to do it again, like 1-3 lines of text

I know Theo hates Google AI studio, but one of the things that I really like about it is the ability to edit the chat history. If the chat is getting long and has a bunch of context that distracts the AI I can go back and just delete whole messages, edit messages that have some important context but a lot of fluff, or delete early versions of a file generated by the AI. The ability to edit the chat history is something I really miss when using chat agents in IDEs.

One of the big problems with "don't distract the agent" is that's exactly what we have been putting up with from terrible managers for decades - that's why we end up doing exactly the same thing to agents.

Thanks for the great video. I find agents.md extremely useful for large legacy codebases. For example, when you want to introduce a new pattern but don’t want to create a huge PR touching the entire codebase at once, it’s very helpful to tell the model: from now on, we use this new pattern, even though many places still use the old one. The rule can be: if a file is touched, migrate it to the new pattern; otherwise, leave it unchanged. That helps a lot to prevent the model from constantly getting confused by old legacy code.

The system prompt doesn't technically "have precedence" over user input. All tokens, including that system prompt part, is treated the *exact* same way by the transformer. In practice yes, the text in the system prompt does carry more weight than if the user had written it but that's only because the model has been trained on that system prompt and has adjusted its internal weights/bias towards the sentiments in that prompt. If you're sufficiently persuasive in your user input then it can overpower the system prompt.

Thanks for accessing academic papers. People work really hard to get into the weeds of validating and benchmarking in a scientific manner. As a fellow AI researcher, we appreciate it!

I love how programming went from logical followable logic to magic text creates magic text that we dont know how a magic text will influence the next magic text, yeah vibes amirite gyze~~~

I miss when computers were deterministic

Another reason why looping through tasks from bash works like a charm.