i started using claude in a new way and i've already pushed out like 6-7 PRs this morning so far. this is way more than i would've done previously
Timeline
Post
Remote status
Replies
27
@shibao let me in on your secrets
@shibao i'm getting so much done i'm tired every day
@RustyCrab add this to ~/.claude/commands/bootstrap.md` and type in `/bootstrap this is a problem` whenever you have to do anything manually. the goal is to do as little as you need to, to get quality output
@lain i was already tired every day before ai! actually integrating ai into my workflow has been a huge improvement for this, i was borderline about to burn out completely until ai saved me at work tbh
@shibao yeah, there's a lot more productivity and much less "how do i actually call this api do i really have to figure this out for 3 hours" type stuff
@shibao you're one "this is a game changer" away from this reading like a linkedin post
@subnetter @lain all the ones i tried out were shit
@subnetter @shibao yes, but for programming they are not quite there yet. for other stuff they are quite usable, gemma 4 is really good.
@lain @shibao I've been focused on programming lately, and with Claude and Gemini I've been able to learn stuff and write rust code fairly fast as a programmer, but I don't vibecode, I just ask about concepts or how-to-do-thing and I then implement. I also did a clean room thingy, I had vibecoded a script for something I do, and now I'm implementing that on my program, so I told it to give me the details as if we were clean room implementing smth. Very fun.
For other things like fixing bugs in build stuff (nix, c++, gentoo [in the past]), it's also really good because I don't have to get fucked and either cope with the failure or wait for a fix.
For other things like fixing bugs in build stuff (nix, c++, gentoo [in the past]), it's also really good because I don't have to get fucked and either cope with the failure or wait for a fix.
@lain @shibao i had thought of this and "hmm will i lose the ability to find out stuff for myself" and the answer I think is, "no". Instead of having to get pissed at trying to find how to do something, or read a lotta docs, Claude/Gemini can simply point how a stuff is/works and then I do the work myself. Hell, this reminds me of trying to navigate the openAI api endpoints and strugglling to find the exact values for the thingies.
I've found that I am forced to develop an understanding of libraries I use anyway because the models (GLM 5.1 recently) get stuck on the implementation details and I have to ask very precise questions. I gain this understanding faster because I don't have to dig through the sources myself, but in the end I still know now how this rendering library packs the color channels and alpha bits. If I didn't ask, I would have just been stuck.
might be a skill issue
might be a skill issue
@WandererUber @shibao i don't see this behavior but i also have no idea what you might be 'doing wrong'. maybe i should just record a session of writing some tool with it.
I would love to see that
@mischievoustomato @shibao yeah i've been fixing / modifying software a lot. it's really the age of the individual's software.
@lain @shibao yep. I haven't modified much currently, but like, sometimes stuff will break for me on nixOS, and there might be a PR waiting, there might not be, and with Claude I can fix that for myself. Also giving it a nix derivation and an arch pkgbuild and telling it to update the nix deriv according to the pkgbuild and it doing great. It's really good.
@lain @MacGuffin @shibao ha, i know how2program a bit, but i dont have money for anything. I push the limits of what my laptop can do, and use the free tier stuff of things.
@mischievoustomato @MacGuffin @shibao here's some free stuff https://nitter.net/modal/status/2041820290716729379#m
@shibao @RustyCrab can you elaborate ? do you literally say "this is a problem" ?
@kon @RustyCrab if you can't just do `/implement xyz` and it will automatically create a PR that fulfills all the problems, then do `/bootstrap where it fell short` wherever it fell short. then it will update scripts or whatever so that hopefully it gets better next time. there will still always need to be some manual intervention because ai models are not that convergent but you can at least get it like 95 percent of the way there eventually
@shibao @RustyCrab @kon i feel like people are afraid to say how good it is. if you use these tools somewhat skillfully, the results will be better than 99.9% of human programmers.
@shibao @lain I heard a few good things when you use vulcan with some of them, but I still refuse to use anything related to OpenAI and Anthropic because of my ethical lines. I've toyed with Qwen3.5 for simple shit like emails and basic documentation lookup.
I am not a swe but none of my co workers have productivity boosts from LLMs and a few have leaked sensitive data on them 🫠
I am not a swe but none of my co workers have productivity boosts from LLMs and a few have leaked sensitive data on them 🫠
@subnetter @lain you kinda have to be a good dev to use them well tbh which is kinda funny
