This weeken I vibe coded a lot
Since at my work there is a huge push to use AI I had to polish my engineering skills with some vibe coding at home.
Usually at work I use Github Copilot with GPT-4o + paid models like Claude code 4.5+.
Since I’m an European I did not want to use proprietary models made in USA to slop code at home. That’s why I decided to use Mistral [1] + Openclaude [2].
Following post shares my views on differences between them.
What I made this weekend
So it was a simple application which should allow users to share their CDs via Post office.
Idea is simple. When one person is registered to my service they receive a share registration link which they can share with others. When this other person joins via referral link they can ask users of my service to share their address and send some CD with good music via regular post office.
Prompts I used
A bit of PHP, Docker and pure HTML.
Create me a simple website which has a purpose of sharing CDs via post office.
The system should work this way:
Whenever you register as a user you have a possibility to invite new user via link. This link allows other to register. When user register it's stored as in a graph bellow invitation. Then such registered user can ask others for address to where they can send CD to. When other user share their address the person should send their CD via post office and share photo of CD in envelope.
After CD is received that person receive karma points.
System requirements:
Use pure HTML without any CSS.
Backend should be written in PHP as REST API.
Front end should connect to backend via REST.
For database use PostgreSQL
Everything should be setup within docker container which should allow for development.
Keep it simple.
Suggest what pages should be implemented and imlement them. Start with front-end.Since it did not work in the end I also used simpler prompt based on nextjs.
Create me a simple website which has a purpose of sharing CDs via post office.
The system should work this way:
Whenever you register as a user you have a possibility to invite new user via link. This link allows other to register. When user register it's stored as in a graph bellow invitation. Then such registered user can ask others for adress to where they can send CD to. When other user share their address the person should send their CD via post office and share photo of CD in envelope.
After CD is received that person receive karma points.
System requirements:
Use Next js
For database use PostgreSQL
Keep it simple.
Suggest what pages should be implemented and imlement them. Start with front-end.Time I spent approving and aligning prompts
Both sessions took around 1 hour. During that time I had to approve a lot of file creation requests and multiple file edit ones.
Sadly in the end not a single project run reliably. On the surface both codebases looked promising but as soon as I started to run the project did not start properly.
After that I spend quite some time to fix the issues. I tried to debug errors within Openclaude. This did work well. Even thought the first setup was done with Docker and it should run out of the box it did not. At first I thought I could fix those issues myself but when I started to dig into generated code I just gave up.
My observations between Claude, GPT4 and Mistral
I see a huge difference between Claude 4.5+ and GPT4. Same between Mistral and GPT4. But when it comes to Claude and Mistral Code Latest there is no difference. Both generate unmaintainable code which from surface looks really promising. Without a lot of tweaking or aligning prompts it’s not possible to generate complex application on first try.
Observations between Github Copilot and Openclaude
Here there are many differences. First of all Github Copilot is less autonomous. Even though you can have agent mode which will generate you a lot of slop when it comes to autonomous behavior Openclaude is just way better.
You can specify whatever you want and it’s figuring out what to do right away. As soon as first big issue starts both fail. Especially when they have to fix a lot of generated code.
Is programming dead?
So far without involvement of senior developer both models fail to deliver. As soon as your vibe coded software starts to break the amount of tokens LLMs consume to fix it is just enormous. Agents start to read data from many files wrangle with it without success. Whole process is slow and after some point even if the issue is fixed the amount of code increases.
But that’s not the worst. The biggest issue I have is that generated code is hard to maintain. There is a lot of it and without proper understanding of technology used by LLM there is just no way to jump into it. Even if you ask LLMs to generate you documentation. Things get hard to grasp pretty quickly. Especially when you have to maintain this codebase later.
What’s next?
I think following days I will try to get both projects up and running. This project is my goal for this month as I have a lot of CDs that I would like to give people for free :D
But so far I don’t see possibility to fully vibe code it. Even at work when I slop code it’s usually specific, surgical, precise change for some part of code.
Sources
[1] “Mistral website.” [Online]. Available: https://mistral.ai/.
[2] “Openclaude website.” [Online]. Available: github.com/Gitlawb/openclaude.