Somewhere in the last six months I stopped opening my editor on a Sunday, and I don’t mean the work one, I mean the other one, the editor where the projects nobody asked for live.

There was no burnout and no crisis, nothing you could point at and name, which is exactly why it took me so long to notice at all. Work was good and busy in the way that good work is busy, reviews went fine, features landed, and every metric that measures a software engineer from the outside was pointing up and to the right, so this isn’t a story about a job or a team or a workload, and on that front I’ve been lucky and I still am. Nothing was wrong, something was just gone.

The symptom I ignored

Orgsidian (one of my side projects) sat untouched for months, which on its own means nothing, because side projects die all the time and dying is most of what side projects actually do. What was different this time is that I never decided to stop, because there was no evening where I sat in front of it and thought “this isn’t going anywhere”, I simply didn’t open it, first for a couple of weeks and then for a great deal longer than that.

And when I finally did open it again, the first thing I felt wasn’t curiosity about where I’d left off, it was the urge to catch up, to drag the thing back into a state where it was worth having, which isn’t the same as wanting to build something and is really just wanting to close a gap.

I told myself I was tired, and maybe it really was just a full stretch of life with more responsibilities than I used to have and less spare attention to spread around them, and reasonable explanations like that were available so I took them, because they were true enough to hold.

Shipping more, enjoying it less

I want to be careful here, because the lazy version of this story is a complaint about AI making the code worse, and that isn’t the story I have. The code was fine, most of it is still running, and if you’d asked me during that year whether the tooling was worth it I’d have said obviously yes, and I’d have been right.

What I remember about writing it is nothing. There’s no afternoon attached to any of it, no wrong turn I took first, no moment where it clicked, not one detail I could tell you now without going back to look it up, so it exists and my name is on the commits and it left no trace in me whatsoever.

That kept happening, and it wasn’t only happening on the clock, it happened just as much on the things I was building for myself, which is the part that should have bothered me a lot sooner than it did.

The pleasure was in the session, not the artifact

Here’s what I think I had wrong the whole time: I assumed that what I liked was building things, objects, repos, features, deployed URLs with my name somewhere behind them. I don’t think that was ever really it. What I liked was the two hours.

You know the ones I mean. You have a rough idea and no plan, the obvious approach turns out to be wrong in some interesting way, you end up reading the source of a dependency because the docs won’t tell you the one thing you actually need to know, and then you look up and it’s dark outside and you’ve got maybe twenty lines to show for the entire evening, and you feel great about it.

AI is very good at compressing exactly that, and it’s supposed to be, because optimising for the artifact is the entire product and nobody is hiding it. But when the artifact turns up without the two hours attached to it, you end up with everything you said you wanted and none of the thing you were actually there for.

I wrote about the other half of this a few months ago in AI and the Craft, the half about what you trade away when you let the tool do the thinking for you, and that half is much easier to talk about, because saying you understand your own systems less sounds serious while saying you’re not having fun sounds like whining.

It isn’t whining, though, it’s the reason I have this career in the first place. I didn’t arrive here through a plan, I arrived because at fifteen I found this more interesting than anything else on offer, and that kid is why I’m any good at the job now, and nobody was paying him to be interested. Whatever keeps him around seems worth maintaining on purpose, the same way you’d maintain anything else you depend on.

So I’m starting a project with the AI switched off

Not as a protest but as an experiment, with a single question sitting behind it, which is whether the thing I lost is actually gone or has just been crowded out.

The rules are deliberately dumb, because clever rules turn into a system you end up maintaining instead of a project you actually build.

• No chat and no agents. I’m not disabling autocompletion, because it’s genuinely useful and pretending I’d give it up would turn the whole thing into a performance, but what stops is opening the chat panel to do my thinking for me.
• Docs are fine, reading the source of my dependencies is better than fine and actively encouraged, and searching and forums are fine too.
• No deadline, no roadmap and no launch, because if it never ships the experiment has still run.
• One hour at a time, usually at the weekend, and the session is the unit, not the milestone.

That last one is the whole point, because if the session is the unit then an hour spent reading with nothing committed at the end of it still counts, whereas under the old accounting that hour was simply wasted.

What I expect to go wrong

I haven’t started yet, and this is the last thing I’m writing before I do, so everything below is a guess, and guesses about your own future behaviour are worth roughly what you’d expect them to be worth.

The one I’m most confident about is the reaching. I know that reflex is there because I caught it earlier this year with autocomplete, opening a file and then waiting for the suggestion before I’d finished deciding what I wanted, so I don’t think the first thing I notice will be the slowness, I think it’ll be my hand going for the shortcut the way you reach for your phone in a lift. And it won’t happen on the hard parts either, it’ll happen on the small annoying ones, because the reflex has very little to do with difficulty and a great deal to do with not wanting to sit inside a pause.

Then the slowness, which will be real and which I’m not going to dress up as something it isn’t. Some of it is just dull, because reading a changelog to find one flag isn’t craft, it’s admin, and doing it by hand doesn’t make it noble.

And then the likeliest outcome of all, which is that I do two or three sessions and quit, not dramatically, but in exactly the way Orgsidian ended, by not being opened. If that’s what happens I’ll write that one down too, which is a much easier promise to make now, while I still think it’s unlikely.

What I’m not claiming

This isn’t an argument, it’s one person and one project that hasn’t even started yet.

I haven’t stopped using AI, I use it at work every day and I do it on purpose and I’d defend that, because refusing a tool that demonstrably works just because using it feels bad is superstition, and shipping is still the job.

I’m not telling you to try this either. There’s a version of this post that ends in five bullet points about reclaiming your craft and I’d rather not write it, partly because I don’t know that any of this generalises and partly because I don’t know yet whether it’ll work on me.

I’m also fairly suspicious of the nostalgia running through all of this, because it’s very easy to mix up “coding was more fun before AI” with “coding was more fun when I was younger and had fewer responsibilities”, and some of what I’ve described here is probably the second one wearing the costume of the first, and since I can’t separate them cleanly I’m not going to pretend that I can.

What I’ll say instead is narrower than all of that. What I lost wasn’t skill and it wasn’t time, it was the part of the work that happens while you’re doing it, and that turns out to be the part I was in it for.

The trap I can already see

There’s one failure mode worth naming now, while it’s still theoretical, and it’s that the pleasure quietly relocates, so that six months from now the good feeling starts arriving from writing about the sessions rather than from the sessions themselves, which would be the same substitution I’ve just spent an entire post describing, one level up, except that this time the artifact is the post.

From the outside that would look identical to this going well, with output and consistency and a growing body of work to point at, and you’d only ever catch it from the inside, and probably late.

So, for the record, the sessions are the project, and this is a note about them written before the first one.