Such a helpful prompt:

"explain ... in the context of ... to me in 3 steps:
1) high level
2) specifics
3) nitty gritty details"

Entering the world of hooks and punches #marketing

The Apple Oracle has accepted our submission.
Jesus fucking christ.

I now also asked Cursor ( of course )

It suggests that I delete it's very own .cursor file 😅 🤞

@garriguv yes - we'll likely go with Supabase. @yvg is on it already :)

Wow - great Cursor bug. The old change context of past chats is still being shown alongside the even older code that existed before. E.g if I visit a file that I've changed 2 days ago, it's full of change marks... (Yes, pressed "Keep all" earlier)
How can I flush this? #cursor

Bummer! Apparently, Google Firestore doesn't support full-text search. I guess it's time to look for a better DB.

Planning next moves

TIL: it's incredible easy to vibe-code a migration script today that triggers MILLIONS of reads because of a N+1 query pattern 😱

We're running our backend on Firebase Cloud Functions. In a way it's great to (re-)learn about the DevOps/SRE stuff from back in the days with gcloud again.

On the other: Man, it's easy to get lost between firebase auth, gcloud auth, setting the right quota project and user plus the permission madness that is necessary to do anything.

i tried vibe-coding a swift app on the weekend. This is a whole different level of a pain. I'd say, we're not quite there yet #AI

wow - after what feels like 10 years I've opened Twitter again. It's an absolute wasteland 😭

I was playing around with the Vertex AI prompt optimizer notebook just to find out that:

- I need a custom metric
- in order to use custom metrics, I need to create a Cloud Function (which I did)
- the notebook doesn't even support that.

That's all too complicated! So I coded a super simple prompt evaluation loop with Cursor in 2h to measure LLM results using our system prompt against a ground truth dataset.

4 iterations later, performance improved from ~ 70% to 96% !! #PromptEngineering ftW

Built fluado's MVP with AI.

Vibe-coded a React prototype in 12hrs. Chaos ensued. @yvg saved us with MVVM + strict typing—suddenly fast again.

Then reality hit: payment providers, app stores, legal docs. That's what actually kills velocity, not the code.