Is Tilck written using AI?

[vvaltchev]

I saw that it has a CLAUDE.md file, does it mean that this project is AI slop or contains a fair amount of AI slop?

[vvaltchev]

Absolutely NOT. Tilck was created in 2016 and there is a full history in the commits. You can see every mistake, from the early experiments to what's today. However, the tech world is moving fast and I can, on average, spend only a few hours per week on this project, and not every week. There is a ton of boring stuff that Claude Code can help me with, as it can dig into the code and do quite some advanced dependency analysis, in particular for things like CMake and the toolchain package manager. So, since commit 175b68f (April 2026), I started using very carefully Claude to get some help and move faster to accomplish at least some boring things like CMake code refactoring. I control and review Claude's code at every step and intervene every time I don't like something. That way, the project could continue to grow, at least from the infrastructure point of view, despite me having very little time for it. I'd rather do that than taking months to find the time and energy to do some unpleasant work like making the build to work on a new platform like Darwin.

I think we should all be very careful with AI tools and avoid "vibe coding" 10k lines of code per day, because that will generate a lot of tech debt and we'll completely loose track of how all of that code really works. But, on the other side, with the understanding that:

1. LLMs are machine tools that could help us do the work, not do everything by themselves
2. LLMs can grind and provide analysis on a medium-large codebase that we cannot remember entirely anymore, in particular if we don't work full-time on a certain project.
3. LLMs being a 100% correct is not necessary: it's enough to give us a good hint about how things work: we should evaluate what we actually want to happen and evaluate if code produced by LLMs makes sense or not.
4. LLMs can suggest a bugfix by doing a lot of try and error that we often do manually anyway.
5. LLMs can help a lot in particular when a 3rd party tool doesn't work anymore or doesn't work on a new platform. Neither we nor the LLM know the codebase, so we'd need to spend a long time debugging a 3rd party tool. An LLMs can help us narrow-down the problem if reproducible. It can search online for similar bug reports. It can search in the release notes and determine if a newer version of the tool has already that specific bugfix. It's a lot of analysis, that's sometimes not worth doing manually, as the tool is not a core part of our product.

I believe it makes sense to embrace the new technology and benefit from it.

[clarence112]

I have been following this project for a while. I am deeply, deeply disappointed that AI is now being used. The fact that you are carefully reviewing its output does not excuse its massive, negative real-world impacts.

[cozis]

@clarence112 No point in having hard feelings about it. I agree there are a lot of bad practices going on, but a better way must be found alongside AI, and not by ignoring the world around you. For what it's worth, knowing how @vvaltchev works, I am sure Tilck's code will be as good as ever if not more