Looking at my GitHub (@tomlin7):
- 2021: 2000+ contributions.
- 2026: 4000+ contributions.
Back then me and my friends would get burned out while trying to add new features to our side projects. Stuck, tried and failed.
Also I used to merge halfassed features (which somehow work, but not really). Because you don't always get someone to really point you at the right resource.
Adding an integrated terminal inside my app was one of such features. It took me a year to go around it and finally figure out that it was just PTYs. Similarly I discovered LSP, DAP, Tree-Sitter (a friend told me about it), and more.
...and I have seen those get reverted instead of trying to improve it. Because zero implementations on a codebase is better than halfassed implementations on codebase.
Google sucks when you don't know the jargon. You can't Google for "PTYs" if you don't even know the word exists. You just end up searching "how to run cmd in python gui" and get 10-year-old broken subprocess.Popen hacks.
LLMs are great at reverse-searching concepts:
- Describe what you want in plain words.
- It gives you the actual standard: PTYs, ANSI escapes, AST parsing, IPC sockets.
- Saves months of wandering in the dark.
Back then coding sessions were pain. I used to actually "give up". But now there is no easy "give up". I use LLMs and figure out the RIGHT resources I need to look at before attempting. I can also predict all possible outcomes beforehand from these sources, before putting any efforts on, because in some corner of the world someone has already solved that thing either fully or partially.
Failing on paper instead of in git history:
- "If I pass editor events through this queue, where does it deadlock?"
- "What happens when the output buffer fills up?"
Someone on GitHub hit that exact wall 8 years ago. Finding out the gotchas before writing 500 lines saves weeks of code that just gets reverted.
Also you can make yourself feel like you are building for Stark Industries, if you manage to set it up really well.
One of the codebases I have that AI has barely contributed to is Biscuit, and it's mostly:
- unit tests and their inputs
- fixing conflicts
- configuration files
- some actual features