Anyone have any experience doing that?
I have a few ideas I’d love to try for my business.
Maybe my opinion is controversial for other CS people, but I like vibe coding. I do “serious coding” for my job, and then have dabbled with some vibe coding for side projects that I was always procrastinating about. It feels quite relaxing.
If you are planning on using it for business, there are two concerns:
- Security
- How to connect your application with other services and databases. Vibe coding usually doesn’t cover this
I’ve been writing 100% of my code with Claude Code for almost a year now, but I review everything it produces and keep it on a pretty tight leash.
Things I think are really important for good results:
- Use a paid plan with a generous limit. Context is everything for LLMs, and the free/cheap plans have to limit it so much that it it gets lost very quickly on anything non-trivial. I have the $100/mo Claude Max plan and I’m really happy with it.
- Use a proper harness - Claude Code, Cursor, whatever
- Use a spec-driven-development tool like OpenSpec or SpecKit. It’ll write the spec for you, asking you questions as it goes, and then you can review the spec (without being a programmer), clear the context, tell it to implement the spec, and usually get pretty good results. You can spec a bunch of stuff at once if you like, but implement one feature at a time.
- If you’re not reviewing all the code yourself, then get it to review its own code (in a clean context) at each step. And/or ask it to review the whole project for refactoring and cleanup opportunities from time to time.
- To save a ridiculous amount of time agreeing to let it do things, run it in a safe environment (like a VPS) and disable permission checks. Make sure it doesn’t have access to anything important (like your production environment) from there. DON’T do this on your normal computer.
It’s fine but vibe coding will probably die in the next 1-2 because it’s too expensive
Expensive due:
-
high token consumption / compute requirements. LLM providers are incentivized to create unnecessary long code, because the LLM provider makes more money from longer responses. This consumes the user’s daily and weekly token limits. A team of 3 agents can consume a 5 hour session budget within 30-60 minutes. Non-monthly subscription API pricing solves this problem, but exceeds the budget of most vibe coders.
-
more human time required to review the longer code. Vibe coders usually don’t review code, but larger companies with revenue-generating code cannot blindly accept code from the LLM. Previously there’a maybe 100-1,000 lines of code to be reviewed per day per team of 4-6 engineers. With LLMs, it’s closer to 10,000+ per team. Everyone creating random shit and spamming the company codebase, just as they spam internet/social media
In the past 1 month, many companies realized LLMs cannot actually replace their employees and a new hiring spree is starting. Need to re-hire all the people who were fired.
I would say try Google’s antigravity or Antropic’s Claude Code.
You just need to be good at describing what you want and how you want things to change.
My Taiwan Flag Design website is mostly vibe coded using antigravity. I am not good at designing user interfaces, and I think LLMs are pretty good with coming up with a basic but clean deisgn.
From what I understand is it’s hard to update the application later without breaking things.
I hear it’s less of an issue with Claude Code. It has been my experience with Antigravity, but nothing that git or describing wheat went wrong and how you want it to work couldn’t fix.
Not actually sure how well it would work for a non-developer – but I’d say close to 95% of my code is LLM generated at the moment (I do review most of it, and pay extra attention to the security boundaries though). Opus 4.7 genuinely is pretty good though, and I don’t have to steer it nearly as much as I did 6 months ago. I’d recommend a $100 claude code subscription, and a $20 codex (chatgpt) subscription to get you started.
- I always have it generate a plan .md first, and usually do have a few suggestions on it (you’d probably care less about model and api shape – but you could maybe keep it “spec” oriented for similar results)
- I then have codex review it for issues and improvements (I literally tell claude to call it’s codex cli to do the review, and to push back on anything it disagrees with)
- I then have it implement it, and importantly, ask claude to verify itself. For a website, it can use a browser to spot check itself. For an api, it can call it.
- Again, have codex do a round of reviews
- Always have it generate unit tests as you go
Keep context below 250k tokens if you can – that seems to be the sweet spot (even though they let you use up to 1m). Clears work better, but compact at least after each iteration if you’re implementing something big.
I guess non developers can hack something together pretty easily but anything complex or long term or handling confidential data or payments is going to need some expertise.
But I’m interested to see what Andrew can come up with .
I’ve coded Android apps and game mods. I always use Gemini Pro. It’s surprising how much it knows about very niche game APIs and such. But sometimes you have to go round and round for two hours to fix something and it turns out it was a simple change because it was hallucinating.
It really helps to start a new chat once it gets long and you’re changing the goal. The longer the context the worse it seems to get.
It’s also annoying to have to copy/paste code back and forth, but I haven’t tried anything like Codex etc. I also have to always tell it not to change unrelated code, not to strip out my comments, etc etc, to get a clean diff.
Great tips! I would add using Caveman mode to save on tokens (though ClaudeCode’s new limits are pretty good): JuliusBrussee/caveman: 
why use many token when few token do trick — Claude Code skill that cuts 65% of tokens by talking like caveman
That’s why Google created Antigravity.
Well but then I’m not in the IDE purpose-built for what I’m doing, eg Android apps. Also VS Code sucks.
Ok guys. Any advice for me as a first timer?
My idea is this after thinking of the simplest app I can build that adds the most value to my business.
I sell professional products to beauty professionals. Some of them relate to working with lashes and brows and requires precise timing. But since you only can work one side of your lashes or brows at a time, you need 2 timers.
What most people do is use their phone timer and it’s fine. But it’s kind of annoying to have to set 2 timers every time and sometimes people forget which side they did first if they’re just starting out learning and don’t have a consistent work flow. So the app would have 2 timers on the same page. Left eye timer and right eye timer.
This app has 0 monetizing value imo but it’s only for us to put our company branding on and have a button to our shop. It’s not meant to be revolutionary, there’s no VC money I’m pitching for, I’m not looking to exit 10x revenue at some point.
And tbh someone can easily clone the idea and there’s similar apps for other things like fitness and cooking so in some way I guess it’s a positive that this concept exists in the App Store already in other areas.
Our advantage is we have a significant following on social media and tens of thousands in our mailing list so we have the ability to push the app out there to the community.
Some people might see it as unnecessary and that’s fine. It’s more a quality of life preference and brand prestige type of thing to say we’ve built this. it’s free so there’s no commitment.
I need this to be a mobile app for ease. One possible complication is this app needs to be able to keep time in the background and notify the user with some sort of alarm when it’s done even if they’re doing something else on the phone.
I’m doing a lot of research of how best to approach this with vibe coding. I learned python for finance but I barely remember. I can read basic code and understand what I’m looking for on my websites backend and use AI to add some very basic coding sometimes. That my level of coding.
I would love any advice from people who may have some experience and insight
2M now, and there’s a new /autocompact feature which lets you define a limit for it to “aim” for, as well as a simple “auto” option where it will decide when to compact based on the relevance of the current context’s content vs the next steps to be done in the work. (This works best when combined with planning.)
Gemini does really well for things like deep research, but the code is really quite bad, like night and day compared to Claude.
The mobile space is in many ways completely different from web apps, in other ways exactly the same.
There’s two main categories of mobile apps:
- “native” apps (i.e. run locally, can be used offline, etc)
- apps that are basically wrappers for websites
The former is more difficult, but probably still better suited for your needs.
Also, be aware that the vast majority of apps are monetized via the info they gather, which is everything from the usage of the phone (e.g. call durations, screen time, etc), when the person is walking vs driving, GPS coordinates, what apps are loaded and/or active and when, etc, etc.
Not saying you should go that route, but if you’ve got 10k+ users it would be an option, and a lot of app-building frameworks kind of assume you’re going to do so, (which is why I’m mentioning it, because you’ll have to make decisions on that stuff sooner or later as you run into supporting elements for it in the frameworks).
Also be aware that virtually all of the free options for vibe coding a mobile app simply put their own data collection into it without telling you. 
I think a major advantage for us is we do not need to monetarize it directly or with selling data as revenue stream. The revenue stream would be enticing beauty professionals to enter into our ecosystem. I sell products, and I sell a lot of them and want to sell more of them.
But it’s good to note to consider dating collection from free options being a concern to think about.
That would be Tinder…
True. Prior to trying out Antigravity, I’ve been on my trusty vim setup and copied and pasted code from gemini pro just like you. However, hard to beat telling AI to do all the unit tests itself…
You can get Antigravity to write, and then test out your app directly on a phone through adb.