im at the point in my #vibecoding project that every new addition breaks everything i've already created and need to spend another $100 in tokens to fix it. cool.
Discussion
Just use cursor man only $20/mo unlimited
i dunno man. i think im done for a while.
I'm not sure what your approach is. I get the best results by only letting the AI work on one function or class at a time. "Edit the codebase to add this feature" type prompts always lead to trouble.
1. Goals of program.
2. Let's design a skeleton of the program, what functions will we need to hit those goals?
3. Now we write fully isolated function by function. Make sure you understand and document each function thoroughly when it is written.
4. Test and integrate each function as you go.
You also iterate over that for each feature add. A new feature is a change to the program goals, which updates your skeleton, which causes function edits and new functions, which get tested and added 1 by 1.
LLMs love to be super helpful and run ahead. Force them to slow down to your pace. If you lose sight of where they went you can't effectively prompt any more.
Not a Dev and never was, but I have vibe coded projects in production now. One of my vibe code jobs has completely replaced a prior version written by a dev because it is more stable, updateable, and flexible.
The model in my head is I am now the project manager walking a junior dev through a project they cannot grasp the big picture of.
Why? Honest question. My go to right now is Kagi because I can bounce between models at will from a single subscription.
meanwhile I can get 1000x more done by getting into the flow state instead of fighting with the ai to fix things. Vibecoding is great if you’re setting up the initial project but is pretty terrible beyond that.
I was using replit and paying over charges, but switched to cursor and find it a lot better, cheaper and easier to test with.
Damn bro. Sorry to hear you're having problems. 🫂
what i should have done is make commits after i had it 90% working the way i wanted. now it's probably 10% working and deviating more with each attempt to get it to work. im about to just say fuck it, cut my losses and go back to doing what i know how to do 😉
I get that. Part of me wants to tell you don't give up, but then the other part thinks that you shouldn't throw good money after bad.
But keep in mind, I'm a tech retard who doesn't understand it the way a lot of you do. 🤷
nevent1qqspt5d70lqt9ndu3cp59pvtmc0q59jhk49l4xdemgv8tqmsuea5uzcpzemhxw309ucnjv3wxymrst338qhrww3hxumnwhln3m9
Gotta keep these projects small. Lots of small projects that talk to alone another as opposed to one big one.
i added hashtags earlier this morning and the entire app is completely fucked now. im about to rage lol
💯 LLMs do small things perfectly & do complex things in ways that make the most imaginative humans feel almost typical.
There's value in that kind of crazy, but... baby steps.
It turned my kind 20 app into a feed of only reposts and removed every feature. And when trying to add them back, it kept removing and changing everything. It was going so well for so long. Two weeks immensely successful queries and output.
You're still talking to it like a project manager.
Assume you're the CEO and just change your vibe! 😏😅
Seriously, before you ask for more code, ask the AI to help you figure **this high level management/organization problem** and let it tell you what it needs to hear to fix it for you 😉
Don't #vibecode before you master project management 😎
FWIW, As a software engineer with 2 decades experience working on “name brand” Silicon Valley product, you have to be able to tell the LLM when they are being obtuse or using the wrong technique. It happens all the time.
right. and i don't know what any of that is. for example, it tried for the longest time to implement NWC. eventually i gave up. im sure i could have helped it along if i understood anything from the NDK. but since i do not, it trying several methods to move forward just made most things completely broken. i was able to unfuck that, eventually. this latest issue, i can't undo it at this point. trying to add searchable hashtags just completely broke everything and changed the entire functionality of the app to where the app is trying to load every kind 1 note with an embedded kind 20 note and only reposts. it's spaghetti at this point.
Yeah, even I have to carefully guide Claude as it sometimes creates its own algorithm when I specified using an existing library. It’s super annoying
💯
I know some people have a lot of success vibe coding but I'm convinced it's a psyop. I haven't really ever been able for it to do anything useful for me. I find things like copilot 1000x more useful.
i had a 90% working nostr image app with everything but NWC working and now it only shows reposts and kind 1 content and will not revert the changes.
NIP 46 is a lovecraftian recursive celtic knot
I couldn't get that working and removed it once already, breaking and having to fix everything. This is how my bill went from $50-$300 🤣 searchable hashtag feeds did me in this time. Something with NIP-19 encoding and NDK kept popping up as issues refactoring things.
I have found out from my experience that vibe coding works best when you implement all the features you want in one go from the onset. If it is a project you plan on maintaining, after the initial run, vibe coding becomes more of an hindrance.
i would agree. i added most features in the beginning and worked through the plan to implement all features. only after i was done and started adding a couple additional ones did things really start to go sideways and get to an unrecoverable point.
What AI were you using?
Goose and Gemini 2.5 Pro. It was amazing the past two weeks. 😩
"Lemme vibe code a fix real quick, how many tokens could it be?" I feel that 🫂 the burnout from frustration, and burned tokens. Hope you get back to it, even with regular ClaudeAI or the other tools.
For stuff like that, best bet is to write, or have it write three architectural documents, like a rolling window of focus at different levels
1) big goal - with separated milestones, a todo list with checkmarks
2) immediate tasks - elaborate on the current milestone with todos and checkmarks
3) looking ahead, so the AI knows what it needs to anticipate
it was a good learning experience. though im done for a long while. this has been distracting me from nostr:npub195t386gf4f67kg58jw6u6jf8vf9xm5ultrzpgm6em3tr8t046l5q6rgwhq and grownostr.org anyways. i have more important things to work on 😉 shiny objects are fun, but ill let real devs work on apps and ill work on what im good at doing!
Have you built up your test suite before vibing into Valhalla? In my experience if there’s sufficient code coverage it can keep the AI in bounds better. Also try having it stub out new functions, objects, packages, etc and writing tests for those (ie TDD) before trying to implement.
Finally, if you’re using something like Claude Code then be sure to add in tips and guides and basic scripts into your Claude.md so the AI has more context.
See this for more tips: https://www.anthropic.com/engineering/claude-code-best-practices
Thought you were figuring out mcp with goose or something, what happened?
goose was going well, until one thing blew up everything. it's completely unusable and unfixable at this point and not worth the money. oh well.
Unfixable?
When working with an agent the first prompt should be: "create a new commit after each update". So you can easily inspect every diff and rollback.
Oh my bad, i thought you were using goose with ollama. Things take longer but it probably makes you write better prompts rather than spamming to see what works.
Not so cheap
There is a price to pay for skipping and outsourcing engineering.
Always better to learn the skills to fix it when in a pinch. Takes patience and stamina 😁