I’ve done ~10 rides with it so far. Hoping I can convince my wife to use it and save myself $50 a month. That would be my most successful side project by a wide margin.
There are probably a lot of examples like this. Vibe coded software people made for themselves, and other people could use it if they wanted.
All of the code was reviewed by myself, and I’m a programmer, so not sure if that fits the description. I didn’t go through it with a fine-toothed comb, however, and 90% of the review was on my phone. I also did some non-vibed setup for hosting, db, email, etcetera.
To me vibe coding is not looking at any of the code at all, but the definition reads a little loose to me these days, especially on HN, as: did an LLM “type” most of the code or did you? Either way I don’t think the term or definition is a big deal and probably not worth splitting hairs over.
I've lately seen a trend where people use their phone to instruct agent on the cloud to build applications. I presume this is beyond having RDP on your phone
I try to keep my instructions pretty targeted so there aren’t too many LOC and files changed. That’s the goal at least.
I deploy on Vercel, so I get a preview build link I can click in the PR and try the site before I merge to main (trunk based). I use Supabase so if there is a DB migration I just run it in the Supabase console by copy pasting the SQL.
When I merge the PR it deploys to prod. This works pretty well for me.
Database migrations and anything related to calculations have had a fair bit of hand holding. Beyond tests it writes I do still test by hand for confidence.
It’s coming up to a year of use. Claude Code credits has still not exceeded the cost of a paid product. I don’t count my time here because this doubles as keeping my technical side busy, and it’s been enjoyable.
I also built a solution for myself that was largely vibe coded. The underlying schema for inventories, batches, orders, cultures, etc was done in advance to help guide Claude along with documentation, but I'd guess 75 percent of the code is pure Claude.
It has worked really well for a while now. Since it's just me using it and I'm able to roll with issues it causes or verify outputs on the fly if I want to, it's totally fine not being super polished. It meaningfully increases my productivity by allowing me to manage things in a way that fits my mental model and business.
Like you, the cost of the project has been less than a subscription. And the subscriptions wouldn't even do what I needed.
I think the main issue is maintaining sterile conditions but its doable.
I also work at Stripe and will be recommending that we migrate our CPQ off of Salesforce for various reasons (agent force is butt, platform limits are silly in 2025 - 6 meg max heap size for a backend transaction?????).
The "Methodology" that made it work: We moved from initial idea to production in four weekends while I maintained a full-time role. The key was moving past "chaotic vibes" and treating different LLMs like specialized team members: A) Strategic Layer (Gemini Pro): Used for architectural decisions (React/Vite, Node/Express, PostgreSQL/Prisma) and product prioritization.
B) Execution Layer (Claude Code): Used for heavy lifting—implementing the cron jobs, refactoring API patterns, and writing the test suite.
C) TDD as the Guardrail: We never "just coded." Every AI-generated feature followed a strict Test-Driven Development cycle using Vitest. If the tests didn't pass, the code didn't go to production.
The result is a stable system serving 200+ active users with a codebase that doesn't feel "schizophrenic" because we maintained strict cognitive boundaries and context documents for the AI to follow.
My experience so far has been if you possess both deep domain-specific experience and significant coding experience then these coding LLMs, and most notably Opus 4.5, are the greatest productivity booster in the world.
https://aldi-prices.lawruk.com/ https://github.com/jimlawruk/aldi-prices
Problem statement: given a start date and a bible book / chapter, produce a reading schedule for the remainder of the bible assuming one will read 3 chapters every day and 2 extra on Sunday.
So assuming an input of "2025-07-06 Genesis 1," the list would read "Saturday, June 6, 2025: Genesis 1 (3 chapters) \n Sunday, June 7, 2025: Genesis 4 (5 chapters)..." etc.
It created the types, data structures, and utility functions required, and even isolated the schedule generation to a function that used all of them ... and that function was busted. It printed the same book and chapter and date every line.
With a little elbow grease I was able to bring it home. Saved me an hour.
It's a website to read and analyze the Rig-veda.
It's not fully "vibe coded" but lots and lots of AI usage.
I was trying to test the theory if it's even possible to release something production grade with vibe coding. Wrote about the experience here https://kau.sh/blog/container-traffic-control/
From my experience, the biggest difference between vibe-coded projects that go somewhere and the ones that don’t isn’t code quality, it’s whether the builder keeps talking to users after the first version. The “vibe” gets you to ship, but iteration discipline is what turns it into something real.
https://blazingbanana.com/work/whistle - Whistle, which is a complete free, offline voice transcription app using whisper, available on all platforms, Linux, Mac, Windows (with CUDA builds), Android (and iOS as soon as my dev account goes through, who knew paying Apple £79 would be so hard!). To be honest the packaging part was probably the toughest bit and all the different ways each platform needs to build. - Probably my most "successful" one, at 450+ downloads on the Play store.
https://formait.app/ - Free offline document formatting using LLMs to take a load of unstructured notes and give you a nice PDF output. You can load any GGUF model you throw at it as it's implementing llama.cpp, but uses Phi-4 out the box. It's actually quite a useful combination with Whistle, so thinking of integrating voice to text in at some point. This is available on all platforms (except mobile) with CUDA builds available too.
I've never seen anything like it since the original days of the game "The Island of Dr. Brain" released in the early 90s.
I.e. the problem is a lot of time spent on moving the pieces off-of each other. While this is more pleasent in real-life tactile space, not as much fun when using the computer to have to click-and-drag all the pieces around (of course, sorting them etc, is up to the user, but just some kind of initial "see all the pieces in the space without them overlapping each other to the greatest extent possible depending on the total space avaliable given the current zoom settings" ...
The Shuffle button actually tries to spread the pieces out to cover the current zoom level, but it can still result in some of them being obscured. I'll look into implementing a more even distribution.
spocklet-pomodo.hf.space/
I created it as a single main.go with just a single main dependency (gorilla websockets iirc) and I think It's pretty successfull between me and my friends and I am not thinking of monetizing it ever
There is also https://spocklet-beta-pomodo.hf.space/ which has some more features to make it more user friendly that I got suggestion for so yeah
I made it out of complete frustration and the first prototype was built in <30 minutes but I guess I won't really take credit of it because I am just pleasant that I can now use such a software and perhaps other might too.
I don't know but I am very gloomy about AI mostly but prototyping in domains I don't know too much about to create a "just good enough" for my own use case is the only valid use case I find of it I guess.
I've built products that solve my problems and have released one, Intraview.ai -- it's functional, solves a real problem for me and my customers.
That said, as a business goes, it's not a sensation but it's gone from idea to customers using it in less than 6 months. Is it a VC hit, no -- am I happy with where it is and how fast I'm learning -- absolutely!
"Vibe coding" and "AI-assisted coding" are NOT the same.
There's a spectrum of AI use from none to full (vibe coding).
Claude Code is probably the best known example of a product claimed to be coded with AI-assistance to the point of much of it being autonomous now guided by experts. My experience is that is now the norm for many senior engineers and it's certainly the future. I don't know any that are truly vibe-coded but I would imagine plenty of mobile apps.
I also built a Preview Pane Handler for 10-bit videos.
The installers (WIX) were vibe coded as well.
So was the product website and stripe integration. I created a bespoke license generation system on checkout.
I don’t think I wrote a single line of C++ code although the WIX installers and website did receive minimal manual adjustments.
Started with Claude but then at some point during development Codex got really good so I used only that.
My friend vibe coded the entire app to generate thumbnails for YouTube videos.
And, again, I'm not a coder and only know the absolute basics of programming. This is not something I would have been able to do without AI assistance.
There's also the fact that many programmers working on software today both big and small use AI to one degree or another, maybe not to program the whole thing from scratch, but definitely to help ease the process. It's an invaluable tool.
I'm write a few articles here about tricks that work for me when it comes to AI assisted coding: https://foundinglean.substack.com
Tangents (https://tangents.chat) is an Angular/Nest/Postgres app for thinking-with-LLMs without losing the thread.
- Branch: select any span (user or assistant) and branch it into a tangent thread so the main thread stays coherent.
- Collector: collect spans across messages/threads into curated context, then prompt with it.
- You can inspect a "what the model will see" preview and keep a stored context-assembly manifest.
Vibe-coding aspect: about 600 commits and about 120k LOC (tests included) and I have not handwritten the implementation code. I do write specs/docs/checklists and I run tests/CI like normal.
What made it workable for something larger than a static page:
- Treat the model like a junior dev: explicit requirements plus acceptance criteria, thin slices, one change at a time.
- Keep "project truth" in versioned docs (design system plus interface spec) so the model does not drift.
- Enforce guardrails: types, lint, tests, and a strict definition of "done."
- The bottleneck is not generating code, it is preventing context/spec drift and keeping invariants stable across hundreds of changes.
If you define "vibe coding" as "I never look at the code," I do not think serious production apps fit that. But if you define it as "the LLM writes the code and you steer via specs/tests," it is possible to build something non-trivial.
Happy to answer specifics if anyone cares (workflow, tooling, what breaks first, etc.).
I've got some users and the stuff I can do each time I start doing vibecoding is astounding. Obviously 50% the work is just fixing what the AI didn't understood or imagined too much, but having a good AGENTS.md is key (and patience from me) - so that's why I'm buidling LynxPrompt indeed, for having an easy way to own a good AGENTS.md file for my next projects... and hopefully you too.
I cloned Paddle's NextJS starter kit[1] and incorporated my previous reporting code built with Observable Framework[2].
It actually took longer to get the website (domain, terms, privacy) approved by Paddle and my identity verified by its 3rd party than to vibe code the site with Claude Code.
Is it different for Claude?
I have rebuilt it a few times in agent mode while trying to get pmf. I used about 22B tokens this year
https://github.com/rabfulton/ChatGTK
I'm sure the code can be critisized, but I'm happily using the application I wanted that did not exist having never programmed python in my life.
It’s a suite of tools to navigate the Epstein files—it even made it into the news!
Here’s the HackerNews discussion: https://news.ycombinator.com/item?id=46339600
I'm pretty familiar with the underlying stack, which helped a lot since I knew the pitfalls. But pretty much all of the code is written by an LLM.
It's not commercially successful (it's a side project), but still represents a complete project.
https://github.com/pannous/goo (1% handwritten go extensions)
https://apps.apple.com/au/app/calmshows/id6749471333
ios and web app and openapi spec
https://alexjacobs08.github.io/lobsters-graph/
(i built this in search of a lobste.rs invite if anyone willing and able sees this--email in my bio :)
https://chromewebstore.google.com/detail/ai-slop-canvas/dogg...
I think it's decent complexity for something where I didn't even write one line of code. (all Claude Code)
Its been very decent so far. Time will tell if the PMF is there for the MVP, but thats on the product, not the AI generated code slop.
FYI, this was more of a hobby horse + learning project than an "enterprise SaaS requiring SOC2 compliance." I am basically building a toy. So far, I have learned that you can ship code toys very quickly to test a market demand with an MVP.
Funny also how Loveable and the like are hiring engineers like crazy, yet think engineers are not needed anymore. Why not just vibecode Loveable itself? Oh wait I can tell you why.
Some types of programming benefit more from AI tooling than others. For example, prototyping seems to be the most fruitful area. Also, writing small utilities is much easier, to the extent that a two hour job would now take only a few minutes. That's where you get the multiplier posts from.
But working in a large codebase using proprietary libraries is not a solved problem for AI (yet).
It's just that the average engineer does not spend all of their time on things that can be sped up.
Speeding up 1% of your time by a factor 20 simply does not help very much. But for some roles, I'm sure that a 10% net increase in productivity is realistic.
I didnt join them because I dont really want to do all the work that comes with owning a business like the accounting. mostly the accounting. i also dont particularly want to be maintaining an extra couple of systems at present. there mught be vibe coding currently, but not vibe operations
they should have the thing up by june at their very slow rate of building with lovable, but theyre not people who would ever frequent HN.
It's telling that they will put their own applicants through a dozen rounds of stringent technical interviews, Leetcode exercises, use anti-AI assistance tools and pay their staff $500K or more, all for something they advertise as being easy to vibe code away.