Very recently, I released baler.

That was a productive distraction I went through from my primary goal of building a personal productivity app codenamed Heli, it’s still WIP.

Heli has a React Native frontend, and is local-first. I’m still working on a self-hostable (probably multi-user) Golang based backend.

And no doubt, I’m taking help (a lot of help) from Claude since I am not a good front-end developer.

The problem with taking help from an LLM while building larger projects is managing context.

Once your project gets > 3000 lines, LLMs start giving answers that are technically correct but don’t seem relevant to the answer the human is seeking.

It’s also important that one takes control of the core app logic like dhh mentioned somewhere – you shouldn’t let the LLM drive the project, you should know what each piece of code does, even through you didn’t write it.

Even in my case, I had to constantly make sure that my local directory and Claude’s context were in sync.

As the number of files grew, it took me longer and longer to sync these files. The options I already had on GitHub to convert a repository into text didn’t help as the text was too long for me. And I wanted a tool that could also convert a text file back to the source code.

That’s why I built this tool, and have been using it to add more features to “heli”. So far, the experience has been good, and it only takes me 20 seconds to update my whole context with Claude unlike the 5 minutes it used to take me before.

I learnt that if you post a single file to Claude, it’ll take more from your project memory than if you’d split that single file into 5.

Regardless, “baler” does help and I’ll continue using it until Claude releases a “Co-pilot” like something.