I see a future in jj

Since JJ has technically git compat, I think there's 2 things needed for it to take off

1. A good vscode extension (there's two so-so ones that I'm not sure are being updated) 2. LLM knowledge. I ask gpt-5 about doing something in jj the other day, it didn't even recognize it at first. When I reminded it it was a vcs it hallucinated half the commands. I ended up figuring it out myself from the docs

I do think that GUI stuff is important, there's a lot of folks who have said "not interested unless there's a magit equivalent" which I totally understand. I'm not a big GUI person for VCS myself, so I don't feel the same pain.

For LLMs, yeah this is hard for any new project. I use Claude Code, and it does a decent job with jj, it only tries to do git stuff sometimes. I haven't asked ChatGPT about it though, I'll have to experiment with it myself.

You're kind of right about LLM knowledge but the implication is funny. We just cant be expected to learn new things without LLMs...

What a terrible world we're barrelling towards when "LLMs don't know about it" is considered a blocker towards something taking off. Read a blog post, read a man page (which GP says was indeed their solution), or just play with the dang tool, I assure you these things are still possible without the assistance of a predictive text engine helping you!

Since they guy behind it works at Google, I wonder if gemini is better at it.

I don't think Google is best though of as a singular entity, I'd very seriously doubt that if it is, it's because of this.

One problem is that Stack Overflow isn’t being updated with questions about new tools because it’s been crowded out by LLMs, so the LLMs can’t train on these nonexistent new questions/answers.

Doubt the gemini team handpicks training material based on wether the author works at google or not. Doubt they handpick it at all.

> LLM knowledge

Give it time. Once JJ resources enter the dragnet of future model datasets, LLMs will get better at regurgitating it.

> I ask gpt-5 about doing something in jj the other day, it didn't even recognize it at first.

Using VC commands an LLM generates for you sounds like a pretty terrible idea anyway. What if they delete your data? Why not spend 5 minutes searching for information written by a human?

jj undo :)

Both Claude and Gemini can be integrated reasonably by informing them how to use JJ commands or even through MCP

I'm not running them blindly and I do always ask the llm to explain the command since the end goal is not to have the llm do the work but rather for me to learn jj.

+ any good tool will not let you run destructive commands without confirmation unless it's with some force flag and I have enough computer knowledge to not use force flags blindly.

+ "information written by a human" can be unreliable too.