I've built an AI-powered first responder for exceptions in our codebase.

When Sentry logs an error, a new issue lands in our Linear triage queue.

From there, a webhook auto-tags Codegen with a prompt to investigate & act:

- Fixable? Propose a PR
- Noisy or unactionable? Add to ignorelist
- Duplicate? Link and close

Codegen then updates the issue with findings, labels, other relevant issues, and resolution status -- no human touch required.

AI is now our new frontline dev assistant.

cma.xyz. Founder /paragraph, ceo /mirror, previously at Google & Coinbase 

It depends on the task we give it. If it's something that needs substantial context or is unintuitive, it goes down a rabbithole and overbakes things, but for tightly scoped things where we can have a clear prompt, it does very well

Codegen uses Claude Sonnet 4 under the hood by default IIRC

Some other random pointers:
- we have a very long organization / project ruleset in codegen and in Claude.md with pointers to relevant files, how we want things designed, etc. this is helpful
- prompt engineering is significantly important. "do X" will almost always perform worse than "look at files xyz, implement feature with color X and functionality Y, modify backend files z", etc. More specific and tangible is better

Very nice. We’ve got a similar thing with codegen but found it has a bad habit of overbaking things. Have you found the same thing? Any tips to keep it on track?

Finding Claude-sonnet-4 to be most reliable and effective.

Organizing Casts, annotating thoughts, keeping you sane on farcaster.

Hm.  Interesting.  I always appreciate when folks have comms about what they build so we can all learn.

Need to now see what codegen is..

@cstkpr save this