It's not vibe coding it's solution coding.

Euan

• 4 min read
It's not vibe coding it's solution coding.

The "vibe" coding phenomena has taken off with the likes of lovable, cursor and Claude Code appearing on the scene. While I think asking an LLM to do all the coding for you is tantamount to disaster with the security aspects of applications. I have been building on the idea of letting it take my problems and helping me fix them.

“There’s this new kind of coding I call ‘vibe coding’, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.” - Andrej Karpathy

I actually don't like the term vibe coding because to me it is for someone who is trying to make the next big money maker, the next big salesforce or notion. Coding is a skill and personally it needs a human touch. The code is something that is written to tell the computer how and what to do. It is something that programmers have been doing for a long time. The idea of handing it all over to an AI is not something I will cover in my thoughts here but let us follow on this thought of letting the LLM create solutions that can help you. LLMs are clever at understanding code it still needs a human to write it so it can learn it.

I have been working on something using Claude Code which is helping me solve a problem for work. At the end of the day I send a bulleted list of events and items to my boss (my CISO is on the other side of the world) it contains things he needs to be aware of, things that are developing, items that might need his attention during his working day or simply just to give good news.

I have been using OneNote to pull these together and while it is more than capable I thought there must be a nicer way of doing this. So I turned to Claude (not Claude Code) and started brain storming some ideas about how this could be done. The usual appeared:

  1. Email - we all get enough of them already
  2. Teams Message - yes that is how I send it but they got lost in the noise of the day
  3. Sharepoint list - yeah, no
  4. Teams feedback app - it's not feedback

I asked it what about making a simple but custom webapp that I can host myself - at that point it started spewing out artefacts, code, database schemas etc. So I let it go on and on but I moved on to the next thing I had to deal with, but the idea was still there.

Solution Coding

A few weeks later I decided to put some more time into it, I wanted to come up with a few ideas see what Claude Code can really do. Before I set out, I wanted to convince myself that I am not following the hype and following the vibe coding "culture" so I decided to, for some reason, convince myself that I am trying to fix a problem not the next AI infused app that will record your meetings and tell you what you want for dinner.

Where to start, I took my phone and started talking to it about what I wanted this new solution to do for me I spent the next 5 minutes just talking about what I do and what it should have, what it should not do. Giving myself in some respects context. This context to the LLM is key to helping me create what I want.

Context to the LLM is important in many aspects, the idea of talking into the voice app on my phone helped. Uploading this to ChatGPT - I simply asked it to "remove my ramblings" and "create something that will give an LLM context for the task it needs to do". Of course ChatGPT was more than happy to help, which gave me the initial prompt to Claude Code.

Solution coding I think is a much better way of calling this LLM powered coding workflow. AI should be able to help people do things faster not starting from a cold start. I am not planning on making money from this and it is purely to help me. A lot of good phone apps or inventions were first created because the creator had a problem they needed to fix themselves.

“Programs must be written for people to read, and only incidentally for machines to execute.” – Harold Abelson

Where to go from here

I spent an hour playing tango with Claude Code asking it to plan and work out what frameworks and backend to use, to use docker and python, how to host it. I spend most of my quota of tokens on my plan just - planning. No code had been written nothing had moved passed me trying to figure out my requirements:

  1. Needs to be easy to use
  2. Easy to deploy
  3. Python based so I can read the code it is producing
  4. Have a cool name

While number five was the toughest one the others were quite clear. The main thing, I think, between my "solution coding" and "vibe coding" was the planning. I have an image in my head of someone just one-shotting the AI and taking the output as their next big app. For now I am still working though some other brainstorming with many different LLMs. I think I might be over complicating this and in a way trying to prove to myself this is a good idea. I will keep going and return with my process and progress soon.