A place for developers to talk about building on Farcaster.

/dev

I did ~4h of vibecoding at agi house saturday, takeaways:

It's slowwwww, each run is about 2-10m so you only get a handful per hour and they often fail or don't work correctly. Brutal feedback loop so I would recommend doing 3 or so projects at once to stay engaged.

Progress isn't iterative. A few times I ran prompts that improved one dimension but degraded another or broke the app entirely. Since each change is so sweeping, finding yourself out of that mess can be tough or plain slow. A few times I just reverted wholesale, cursor's built in checkpointing super helpful there.

@martinamps and I tried to multiplayer vibecode a bit across two machines and despite the beautiful makeshift rsync script it failed bad. It was cool that cursor could self heal when it detected a file changed mid run from a sync'd change though. Would be cool to have an environment optimized for multiplayer vibecoding.

I think a lot would be solved by faster agent runs (groq?) and limiting scope per run.

Vibe coding: allowing an LLM to write a shitload of broken code that you can’t debug because you didn’t write it

AI agent: allowing an LLM to prompt other LLMs 

MCP: allowing an LLM to do what REST APIs have done forever