My first conversation with Claude was in July 2023. I asked it to look over an NHS proposal for BOE – one of the MySelfCare documents, I think, though the specifics have gone. The model came back with notes that were genuinely useful and a few that weren’t, and I sat there for a moment working out whether what I’d just done had saved me time or merely rearranged it.
Then I closed the tab and barely opened it again for two years.
Not out of any objection. I didn’t have feelings about which AI lab deserved my custom. I just drifted, the way you drift away from a restaurant that was fine but a bit out of the way when there’s a perfectly good one nearer the office. The nearer one was ChatGPT. Everyone I spoke to about AI was using it, the prompts that got passed around were Chat-flavoured, and the integrations and plugins were further along. Network effects work on small one-man design shops in Blackpool the same way they work on anything else.
So the two-year gap wasn’t strategic. It was the absence of a strategy. The work got done, the print jobs went out, sites got built, programmes kept landing on time, and AI sat in the corner of the workflow being mildly helpful in whichever window I happened to have open. I ran some local models alongside – Llama 2, then Llama 3, then various smaller ones on whatever GPU I had at the time. Curiosity projects, mostly. None of them displaced the cloud options for actual work.
What’s strange in retrospect is how invisible the gap is from the outside. The blog master list runs from 2023 to 2026 with no obvious seam. The middle of that span is real – I lived through it, the work happened, the print jobs and the homelab builds and the family business all carried on – but the AI thread through it was someone else’s tool. The posts from that period are written from memory and from the artefacts that survived, not from chat logs I can scroll back through.
The thing that brought me back was Sonnet 4.
That’s the honest answer. There wasn’t a moment of principle or a community I wanted to join. There was a model that had jumped, and I noticed. I opened Claude again for some specific job I can’t now remember – the job isn’t the point – and the session felt different in a way that doesn’t translate well into words. The tool got out of the way. The writing came back sharper than I’d asked for. The pushback was measured. The handling of long context was no longer a thing I had to babysit.
Models improve incrementally and occasionally in jumps. The jump I came back to was real and identifiable in the way only someone who’s been away can spot. If you’d lived through every weekly update from mid-2023 onwards the change would have been a series of small adjustments. Coming back cold, two years of changes hit at once, and the difference was obvious.
The interface had caught up too. Projects, artifacts, file handling, longer context, the ability to keep a body of reference material attached to a working session – all the things I’d grumbled about in 2023 had quietly been built. The product in 2025 wasn’t the product I’d left. That mattered as much as the model.
I’d also be lying if I said the broader picture didn’t push me a little. Watching OpenAI cosy up to the current Trump administration was its own kind of mood-killer. Not a principled exit – I’d already drifted by then – but it made the drift feel like an instinct worth listening to. Your AI tool of choice isn’t a political affiliation, but the company behind it is still a company, and some of them are easier to keep using than others.
The Friday persona work, which I’ve written about before, was an early marker of the comeback. It was the kind of project that needed the underlying model to be able to hold a character properly across a long session, and I’d assumed for two years that we weren’t there yet. Turns out we were. The persona landed because the model could carry it.
Since August 2025 it’s been daily. Multiple projects in the Projects sense, each tuned for a specific job with its own knowledge base. The Garage AI is literally Claude wrapped in a custom personality and pointed at the household. This blog you’re reading is being drafted through a project with knowledge files, voice samples, and a master list of posts to work through. The tool has become structural rather than incidental.
What I’ve landed on is that the choice between AI tools is practical, not tribal. Most of the commentary online treats it like a football allegiance. You’re a Claude person or a Chat person or a Gemini person, and you’re expected to defend your patch. I don’t have a patch. The tool that does the job I need on the day I need it gets the work. Two years ago that was ChatGPT for almost everything. Right now it’s Claude for almost everything. Two years from now it might be something I haven’t heard of yet, and I’ll move again, and it won’t mean anything beyond the fact that the work moved.
The 2023 sessions weren’t wasted. They taught me what to look for when I came back. The gap wasn’t a mistake either. It was just two years of work, done with the tools that fit at the time.
Drifting in, drifting out, drifting back. Goes nowhere, does nothing – except, occasionally, exactly what you needed.