> Not looking forward to that PR, lol!

Well, the whole workshop was about learning how to make agents trustworthy ... so that you can run thousands of agents without having to review all the code they generate. lol.

I do like the part about making agents more trustworthy though!

> give LLMs better verification criteria

Yes! I have developed this nasty (for me) habit of writing a long list of "acceptance criteria" which both claude and codex stock system prompts seem to pick up on well. It slows me down a bit, especially because when I get some highbrow result it's because I forgot to spec something properly in that section, but generally its where the difference is made. That's also where I try to guide plan generation quality but the indirection in that is often making me not like the resulting plan and editing/tuning it.

> forbidding any use of `useEffect` in react

Not looking forward to that PR, lol!

optimism

I haven't tried using it much yet, but it creates verification recipes for features which should get picked up by review runs is my sense.

 describing how important it is to give LLMs better verification criteria. They also build a lot of static analysis tools to constrain the bot's output to less buggy modes, e.g. forbidding any use of 

What are you working on this week?

I haven't tried using it much yet, but it creates verification recipes for features which should get picked up by review runs is my sense.

I watched a workshop from the author [poteto](https://github.com/poteto) describing how important it is to give LLMs better verification criteria. They also build a lot of static analysis tools to constrain the bot's output to less buggy modes, e.g. forbidding any use of `useEffect` in react.