If you’ve read any of my previous posts about AI you might note that I have been less than enthusiastic, even a bit nasty towards it. It’s easy to mock. Around December 2025 something changed. The LLMs available for writing code took a big leap forward and became a lot more useful. Here’s a post about how and what I’m using at the moment.
But first an explanation about what is and what isn’t AI generated. I’m following Simon Willison’s example and if something expresses an opinion and uses words like “I” and has my name on it then I wrote it. AI might proof read stuff for me, write code documentation, product requirements and AI prompts for LLMs but if it’s got my name on and says “I think” then I wrote it. If code has been written by AI it will have the name of the agent or LLM that contributed.
I was very much set against using LLMs in my work at the start, the toolsets were poor and the results in my testing were not very good. I watched the Reddit groups for a few of the LLMs and the posts started to change. There were still the usual “I one-shotted an app and now I’m a millionaire” type posts but people I respect in the industry started to talk about their toolchains and how they were using it which made me take another look. I was compelled by their use of agents and how it helped them do more work rather than how it was replacing them.
I installed Claude-code and signed up for a free account, I have a Python repository that posts graphs of my energy usage to Mastodon, it’s been broken since October and I’ve not had the time nor inclination to fix it. I gave it a simple prompt, something like “Run @daily.sh and debug why none of the graphs are being produced.” and away it chugged. It fixed the graphs, I prompted it to improve the look of the graphs without any hints as to what “improve” meant and it came up with what you see today. As a stretch I prompted it to look for other energy related data in Prometheus and it suggested gas, which I then added, along with some other quality of life settings for the script. I’ve since signed up for a $20 Claude account for use with a website that I run and it’s helped implement a couple of features so far.
I’m also using Copilot and it feels like it’s a little way behind Claude in it’s toolchain, the CLI has only just been released and it’s agentic capabilities seem linked to Github issues which means it doesn’t fit a developers workflow very well. Copilot feels like really good, really really good autocomplete at the moment but it’s CLI might change that.
Another tool that’s very interesting is Google’s Jules. It’s an agent that you attach to your codebase and give it tasks. So far I’ve had it doing a security review of an app of mine. It creates a branch and PR with all it’s changes and then keeps an eye on the PR comments and makes changes based on comments. I need to work more with this because the first thing it did was remove a default for a parameter and then immediately reverted it with “You’re absolutely right …” when I asked why it had removed it. It didn’t tell me why.
Another LLM I’m using is Google’s Gemini hooked up to my Home Assistant instance and I use it to give my TTS a bit more flavour. It currently think’s it’s Michael Caine’s Albert Pennyworth from the Nolan Batman trilogy which is fun. It backfired though when I told it it was Bricktop from Snatch and spewed a stream of obsceneties at someone who’d only left a door open!