"claude please improve performance of my agent and visualize your experimentation thanks"

Accidentally said "hard" instead of "non-trivial" and they kicked me out of SF

Compacting the conversation is the algorithmic equivalent of going for a walk to clear your headspace.

> You are a self improving system. Every time you interact with the user and learn something, write down your insights. Feel free to build and tools that you feel will speed up your work. Lowkey working

LLMs are increasingly great at managing their own environments, including their context and prompts. Alex Zhang you should let the LLM extract insights from the REPL-loop and build custom python functions to parse specific contexts. Should speed up subsequent runs on same domain.

Don’t build static software, build environments. It’s really empowering to see an agent take a differently formatted excel (which would break static software) and just role with it, and optionally rewriting its tools to handle more formats.

Chaotic-neutral software engineering with Claude. > Catch any error, fix the app, and redeploy

"this startup could've been a Codex + filesystem + websearch"

People who worry about AI proliferating low quality slop have never browsed LinkedIn

the lion does not concern itself with the global consequences of critical infrastructure failure

Never not programming my slides anymore.