Claude released 3.7 a few weeks ago and along with it they released a command line programming assistant - Claude Code. This was in limited preview on launch but last week was opened up. I tried it out last night and this post describes the outcome. 

When LLMs came out I started testing by developing a python web app that could use sockets to allow real time collaboration. This was a way of programming web apps that I knew about but had never used. Last year I got a basic app working over a few days with the help of whatever version of GPT was available at the time. It involved a lot of cutting and pasting between GPT and my development environment, but I eventually shipped an app using a technology that I would not have been able to build on my own, with the time available to me. 

Towards the end of last year I refined the app to turn it into something that could be used by my department to submit anonymous questions, and vote on them, for our all hands meetings. This time I used the AI coding editor - Cursor, as well as the command line tool aider. I shipped some new features, and critically used these tools to learn how to write tests that tore down the DB, and created in memory mock data, when running the tests. I got to a working app, but it took a fair amount of back and forth, and there were some bugs with the JavaScript that took quite a few iterations between different coding assistants to iron out. 

I had not looked at the project since last October and went back to it a few nights ago. 

The app has a homepage, a page for submitting questions, a page for voting on them, and a page for viewing questions. (repo - https://github.com/IanMulvany/question-voting-app) There was a stub for an admin panel - I had some ideas about the admin interface, but hadn't implemented them. The page for submitting questions had a clear bug that I'd not fixed. 

I tried out Claude code. 

- It developed the admin page with all the features I needed, in one go. It ran the test suite and added additional tests.  

- It fixed the bug on the submit page, in one shot.  

- It created a much nicer UI, overall.  

In total all of those improvements took about 5 min of Claude time, and about 20 minutes in total. This session cost about $1.50.  

The pricing is a lot higher than using aider, and longer than using one of the tiers of one of the frontier models, but the quality and capability was a lot higher. The effectiveness is a lot higher.  

The UX of the command line utility was outstanding.  

I'm including before and after screenshots below of the different pages in the app. 

Original Submit Page

Claude Code Submit Page:

Original View Questions Page

Claude Code View Questions Page

Original Voting Page

Claude Code Vote Page:

Original Admin Page:

Claude Code Admin Page:

Original Home Page:

Claude Code Home Page:

Final thoughts:

- This is a tiny app, not a production level codebase, but it opens the door to creating things that I just would not be able to do. 
- I still need to have an understanding of quite a lot of the structure behind what I am coding, so this is not a tool for non developers, I need to know what the command line is, how to deploy the code, what git is. 
- I'm struck by the increase in capability of the tool in a very short period of time, and I am certain things are going to get more capable. 
- The rate at which the tooling is improving is a problem in the short terms. It means there is going to be a massively fragmented development experience, and how one goes from these tools working for individuals to working within teams is a critical challenge. 
- The writing of the code is such a small part of software development within organisations, I don't see this as helping. In fact it might hurt a little. It might change the expectations of senior folk like myself about what I might expect of my teams, when the hard things remain alignment, sustainability, creation of customer value. 
- For those looking to prototype or test ideas, this has to be in your toolbox, has to be.