Would be fantastic if Codex could make changes to pull requests that it’s already opened instead of basically starting from scratch when trying to iterate on things

💯

This is it. Using agents means that I’m doing a lot more reviewing/clean up in my IDE and by that time I don’t want any AI “help”

Spent a full day integrating and evaluating Cursor slack integration with background agents. Very positive signals. Led to the first real world case where we’ve had non-technical employees making small contributions to the code base. Absolutely unimaginable 2 months ago.

NBA Finals… Game 7? Never heard of it. 🏆😏

So when does Cursor come out with an iPhone app so I can dictate changes to background agents straight from my phone?

Turns out that even with LLM assistance, working with GitHub actions is still an absolute nightmare

We’ve had a notion of “on call rotation” in engineering for a while. Basically someone who reacts to prod monitors reporting that things are broken. AI feels like it’s going to introduce the “code review rotation” where one engineer is responsible for reviewing all the agent PRs

Increasingly using tool calls as a reliable way to guarantee structured data from LLMs. Want the llm to “select” something? Don’t parse the output. Give it a tool whose only job is to push the selection onto an array that’s captured in the closure.l and just return that when done

Setting up Vertex API keys and I can say with confidence that Google just doesn’t want users. Madness

Awesome workflow for high quality, complex images with nano banana - prompt to create a picture with just the subject. No background. - prompt to add the subject to a background - add more things to background as necessary - pass to flux for outpainting/size correction

AI sdk is great, but I’m not sure of value of just “swapping” models in and out. I’m finding that changing models also means that u have to change a significant percent of prompts in the system to handle new quirks. This is why I favor low level sdks rn

We are shipping LLM based image pipelines at Hungryroot Best pattern for prompt iteration so far: - create a new prompt - apply it 50% of the time - use existing measures to gauge success Increasingly convinced that the best way to verify prompts is just straight in prod

Want: AI bot that can join Google hangouts and code and deploy prototypes live in conversation as they come up

Weird how making your app faster can make your test suite flaky due to your tests assuming your app is slow

AI pipelines are great at getting you 80% and sometimes 85% of the way there but an absolute nightmare for the last mile kind of work. Very good combo to have ai generate something that you can programmatically touch up with traditional automation

Cursor knows about your code base but not your design system. Figma knows about your design system but not your code base. Who’s working on bridging this gap?

JSN is like the receiver version of Earl Thomas. Little undersized. Underrated. Completely irreplaceable