There’s a little explosion of activity on this website and in my Github repository - thanks to Claude.
I’d been pottering about with Claude, Copilot and a couple of other chat interfaces, but had been reading about agents and Claude Code, so plumped for a monthly Pro subscription and started tinkering. And like many others, I’m still open-mouthed about how good it is.
Examples
A couple of examples:
Work needed a design for some error handling in a complex pipeline. With some careful prompts, Claude worked through some design options, discussing pros and cons, then laid out a design, created terraform for the proof of concept application, completed a security review, and created a presentation to convince my team, and the architect, of the appropriate solution. It was an impressive performance, and is now built and part of the product.
Then yesterday I decided to update this website, first time in 18 months. The folder it was in was a mess, I had several copies and couldn’t remember how it all worked. Claude worked it’s way through it, found the real version, dropped the others, and also figured out how I was deploying it.
Use cases
I’m finding it really useful for:
- working through technical designs, exploring options, helping me learn about the problem space and figure out the best solutions
- reactivating old, partially complete projects : documenting them, cleaning them up and getting them to a nice finished state
- helping me develop new projects - I just need to concentrate on the what and the why, leaving Claude to figure out the how.
Drawbacks
I am finding the common problem - it’s easy to build something, harder to deploy it. I’m quite good at forcing myself to finish things though, and I do like documentation - and Claude is making that even easier.
Models and local LLMs
I like Claude; I don’t have much of an opinion on OpenAI’s Sol/Terra/Luna, or Google Gemini or any of the others. They are probably fine - it’s just that the space is changing so fast it’s hard to keep up, so having something that works is great.
I am running Gemma4:e4b on a MacMini with 16Gb RAM. This is very much a work in progress, but I’ve just turned off the thinking part, and that has massively increased performance. I need to figure this out some more.
I’m intrigued by Pi and harnesses - a friend at work has built their own harness to get around the Copilot-is-chat-only limitation and that was inspiring.