Hacker News Reader

The linked post describes being a meat proxy as copypasting tickets into AI and pushing the output without reading. I don't copypaste tickets to AI, I decide on how each task is solved and LLMs only implement what I decided on. I read the code before making a PR. I review other's PR with my own eyes. I never copypaste LLM output to my coworkers in chat. What else is there to it?

View on HN

It's fine. I still program for fun. I switched to Pi as harness, and instructed the model to careful stop early such that I can stare at code, and I tend to give extremly substantive feedback sometimes. It was not easy, and now I hope I found something that clicked for me. I modified the bash tool to force the agent to tell me what he does, and it helps wonders, the agent says, he wants to do X and I can verify if the bash commands do what the agent says. I also use lightweight sandboxing. I am happy that no financial pressures influence my use of AI. What's still open for me is how to discuss it publicly, since there's a large group of programmers jumping at programmers who admit they use AI.

Reply on news.ycombinator.com