2025-12-23
Vibecoding with Claude
I recently bought Claude Max for 140/month and it is quite freeing to use. I have used Claude Code at the 20/month plan and I would hit the cap pretty quickly after an hour or two of intensively using it.
The original reason that I got Claude Max was because I was reverse engineering a binary protocol and I was using Claude. I was making some decent progress but I kept getting stuck waiting for my limits to reset.
This finally got to me as I felt like the reverse engineering was coming to an end and I was hopeful that I needed one big push to get things working. I got claude max and quickly had it finish the job and I was pretty happy with the result.
I then tried dropping it in as a replacement and it failed pretty badly. It seems like Claude had gotten it working for my specific tests and the generalness of the reverse engineering effort was mostly just smoke. I suddenly felt like the ground shifted underneath me as the code claude wrote looked superficially correct but I wasn't even sure how to start debugging it.
I gave up on the reverse engineering effort until I could figure out a better strategy.
This did now mean that I had wasted 140$ and that didn't sit right with me. I felt like I needed to use Claude and so I finally started a project that I've been thinking about for a while.
Technology
Someone had mentioned online that everything self hosted should be just rust and sqlite. This resonated with me as I do think this is basically the simplest possible application with the added caveat that there was going to be some web tech as web tech makes writing applications much easier.
I took the core idea and switched out rust for go. The reasoning was that go is a boring language designed to be written by juniors. It was written to be simple to write and it had a strong standard library. It also had a standard library that included a web server and a templating language. All this combined meant that if I had AI write me something with go, there was a very strong chance that it wouldn't really need much in the way of dependencies.
The only real dependency is sqlite but I think that's the best case as this is an in process database with no database set up involved. This meant that I could get an application very much focused on being easily usable by someone locally.
I also did a Go tutorial and go the basic structure of an application and it made sense to me. I did make some small changes for the structure to match my own aesthethics but overall this gave me an outline that I could have Claude fill in.
Manga Reader
The first project I started Claude on was a manga reader application. I have a ton of manga on my server, some of it is jpegs, some of it is cbz files while others are just regular archives. I have used comic and manga readers before but I had wanted my own version running on my server.
The first thing I did was create a CLAUDE.md file where I laid out the technology to use, the reasoning for why we were using Go and Sqlite. I also laid out the directory structure to use.
I then explained where to find the manga and what I expected.
This was actually an extremely straightforward project and Claude got the gist of it right from the beginning. There was small tweaks I had it make such as adding page gaps and having a single page mode. I also added the ability to zoom in and out as I find that handy for certain manga.
There was changes to make with what it considered the first page of a manga.
Overall, I didn't actually get into the weeds with the manga reader, there is more complexity I could add but I had actually wanted to read something so I got to a usable product and stopped and then got to actually reading.
One feature I really wanted to add was the ability to search nyaa.si for manga and download them. I have transmission set up so this Claude was able to set up a torrent page where I could search for something and add a manga to be downloaded. Once it finished, it would be picked up as available to be read.
This was a great experience with Claude and it made me excited to give it a go with something much bigger.
Video Streamer
With the manga reader working and actually set up. I then got to my big idea. I want to replace Plex. At least on my phone and desktop. The big issue is that I don't know how plex works and I can see that it has broken some of the shows that I watch. Some of them don't appear in Plex even after renaming them and some of the episodes are completely missing.
Plex is fantastic and the interop with the TV is great and I'm sure I could fix these issues but I have created a few different streaming sites for my own personal use. I never did transcoding, I simply re-encoded videos so that they worked in the browser and this was good enough for a single user.
However with Claude, I wanted to see if I could get a real plex replacement going.
Without further ado, it did fantastic. It built a video streaming site using ffmpeg to do transcoding and it works beautifully. I even had it add in the ability to fetch subtitles and it was able to get all of this working in the first day.
I defined the project in a CLAUDE.md and then we were off. It was testing and prodding but there was never a moment where I thought Claude got lost or things were in a broken state. The basic functionality of a video streaming site was done within a few hours and the fact that I could watch an MKV in the browser on my phone was stunning.
The next big thing I had Claude add was the ability to see movies by genre, and have links for actors and actresses. This way I can click a name and see their entire filmography in addition to what I have available.
Similar to the manga reader I also added interop to my transmission server and made it so it could search torrents directly from my video streaming site. This made adding movies much easier.
I also added in a rematch option so that movies could be fixed manually. I also had Claude add in sorting and a download option so I could download movies directly from my site.
I'm a bit blown away by how well the replacement is coming along though I need to actually watch a full movie before I can say that it's actually useful.
The transcoding working as well as it does is the biggest shock, I do wonder if there really is that much training data out there for it as I had thought transcoding was the hardest part. My current plex server uses a macbook with the libraries held elsewhere. This way I can take advantage of a powerful ffmpeg. I think this would work quite well here as well and so I do have plans to move my video streamer to my mac.
The biggest step after building the intial project was adding support for TV shows. This took about 30 minutes of refactoring as I had originally built the application only for movies. Looking back it may have been better to explicitly make TV part of the project. I had it spend another 30 minutes fixing things up as the fetching of metadata seemed to be broken. Otherwise though it is working!
Thoughts
Some quick thoughts are that this something like Claude is going to change the game in regards to self hosting and making little tools for yourself. I don't think any of these applications are worth much in the way of releasing them as products.
After all the AI can build this, all I did was prod it. This is giving me the maker machine from transmetropolitan. A bit more and I won't even need to prod it.
To be fair, I have a feeling these things that I'm building are actually quite well documented and I'm not asking for anything too fancy. I am also deeply worried about the day theres a bug. I know how frustrating it is to use AI and have it go down the wrong path. The reverse engineering project showed my that if the AI loses its way, there isn't much for me to manually do.
Especially since the code is generated automagically anyway, I don't have any interest in trying to fix things. If Claude can't fix it, I might just throw it away and try again.
I would love to see how regular people use this, how much of the ease at which I'm building is because I know what I'm asking for and I'm biasing it into a certain direction. I wonder what a project from a regular person would look like.