Goszczu

What strikes me that all of AI releases show benchmark percentages but none is giving how many tokens used or how long they were running. 80% != 80%. Details matter

I’m astonished how easy was to set up opencode in GH Actions. One command a few clicks and everything works like charm. Now I can chat with big pickle in my PRs

🚀 A week ago, I joined Tonik’s hackathon with my brother, and in just 24h, we built a working PoC for a NeoVim "Live Share" plugin—despite having almost zero experience with nvim plugins. It was super fun.

3️⃣ Rust async gotchas? As Rust newcomers (more async Rust newcomers), we underestimated how much time we'd spend debugging errors like "cannot move XYZ because Foo doesn't implement Send", where Foo was some deeply nested type. On top of that, most of LLMs didn't know how to fix these issues too.

🚀 A week ago, I joined Tonik’s hackathon with my brother, and in just 24h, we built a working PoC for a NeoVim "Live Share" plugin—despite having almost zero experience with nvim plugins. It was super fun.

1️⃣ Lua ↔ Rust communication? After some trial and error, we ended up spawning a separate process and using NeoVim’s remote API for better isolation. We thought about using nvim-oxi or mlua, but after a few tries to set up it correctly, we backtracked to bulletproof rmp.

🚀 A week ago, I joined Tonik’s hackathon with my brother, and in just 24h, we built a working PoC for a NeoVim "Live Share" plugin—despite having almost zero experience with nvim plugins. It was super fun.

I hit a super weird case. TS fails to infer Playwright types, but when I add one empty line to the test body, everything starts to work again.

AI is great and can speed up most of the coding work. At the same when prompting I always feel that it misses context. Often giving suboptimal answers. That's why sometimes I do work manually instead. Is there a way to improve sharing context with AI? I have a few ideas (none of them tested):

Devin style learning via PRs and comments. This is just a broader case of the History+RAG approach. It can work but requires time and effort. I think it's good for big projects but not for small/medium ones.

AI is great and can speed up most of the coding work. At the same when prompting I always feel that it misses context. Often giving suboptimal answers. That's why sometimes I do work manually instead. Is there a way to improve sharing context with AI? I have a few ideas (none of them tested):

A simple AI.md file committed in a repo can be added automatically to each prompt. The question is how it's different from docs or ARCHITECTURE.md. Maybe someone has already tried it.

I really like the Supabase+Postgrest+RLS flow of building applications. But by default, the Supabase Anon Key shouldn't be shared publicly in your apps especially when you want to go fast. 1/n

There are just two important parts to make it work: 1. Proxy all Supabase-related cookies to tRPC. This will make auth and RLS work out of the box. 2. Don't use unstable_httpBatchStreamLink for any routes that can update the Supabase user session.

I really like the Supabase+Postgrest+RLS flow of building applications. But by default, the Supabase Anon Key shouldn't be shared publicly in your apps especially when you want to go fast. 1/n

Did you spot the issue? Users can modify the whole row so nothing prevents them from editing group_id and granting themselves access to the different groups. With the public Anon key in the browser, you are exposing yourself. It's easy to spot these holes by using Mitmproxy or any other tool.

I really like the Supabase+Postgrest+RLS flow of building applications. But by default, the Supabase Anon Key shouldn't be shared publicly in your apps especially when you want to go fast. 1/n

It's crazy easy to shoot yourself in the foot by making a not-so-great RLS policy that will allow easy unauthorized access to your data. Look at this contrived example:

I was frustrated today that there is no easy solution to generate an ARCHITECTURE file from the existing pnpm repo I created a small helper tool. `npx @tonik/arch-gen` or if you want OpenAI magic `npx @tonik/arch-gen --ai` github.com/tonik/arch-gen

Been working on some feature packed boilerplate that will take you a little further in set up than others. One of the features is pre-made e2e fixtures. Once you commit to stack and architecture setting e2e tests becomes very repetitive.