This went so much better than I ever expected. I wanted to install shimeji mascots so bad and no engine I could find was maintained enough or worked for my NixOS so I had claude do a [C]laude [C]lean [R]oom [R]ewrite of shimeji-ee and shimelinux. Even though it’s not perfect yet it lets me have brushbuddy on screen so I’m happy.
Claude did it in 1 prompt. I gave it links to both repos and explained my understanding of how to achieve the process of rendering a mascot onscreen and gave it instructions on how to encapsulate things. I hope in the future I can add wayland support by just extending the code.
Edit: upon reviewing my own sidebar maybe this isn’t a valid post, idk I’m still learning how to maintain communities myself lol what do you guys think?
If you gave Claude the repo, I don’t think that counts as clean room.
I think maybe if you got 1 Claude instance to summarise and document it, and a separate Claude to implement that would be clean.
Well, they claim the code was not used, but I don’t understand why use the project structure then, Readme would’ve been sufficient (I wonder if readme-based counts as clean room):
What was actually used:
- ShimeLinux’s public repository structure and README (what it does, how packs are laid out, what environment variables/config exist) — research, not code reuse.
- The pack format itself — conf/actions.xml, conf/behaviors.xml, the img/ sprite layout, and the attribute schema
They rewrote it in rust, so likely no code reuse agreed.
But “clean room” reversing is more about creating deniability. The engineers in the “clean room” literally never see the original code, so they can’t possibly copy code. The engineers who do read the code never write any code, so not possibly to copy there either. You cant have both readers and writers being the same entity and still call it clean.
You could get Claude to summarise and doco it, and then a different agent to write the new code. Probably open to debate as to whether any LLM generated code could ever be clean though, given they scrape and train on everything. But that is a question for the courts to decide really.
Yeah, my point was that they claim to not look at the code with ai, and only use readme, but I’m not sure it works as clean room
What did you learn working on this?
