The coding workflow is familiar enough: I task Claude Code with moderately complicated work in a codebase—coding, iterating on tests, doing things with the build—then disappear overnight. It works. Not perfectly, but well enough that the failure modes are predictable and nudgeable. I check in, course-correct, leave again. This maps to something recognisable: tech lead with a remote team.

The writing workflow doesn’t have a name. I record thoughts into audio files on dog walks, share them over Tailscale, and Claude Code transcribes and drafts.

Signpost If I’m caught by a flash of inspiration at the gym, I don’t need to find time later. I just talk. The recording replaces the podcast I would have listened to anyway.

It’s not dictation. It’s not editing. It’s closer to thinking out loud with a transcriptionist who also happens to be a competent ghostwriter—but that’s a stretch too. There’s no job title for this. No established pattern I can point at.

And yet it works. I wanted a button to start recording, a button to stop, and a pipeline that shares the file and triggers transcription. Two button presses, everything else behind the scenes. I described this to a model, and twenty minutes later I had it working via Tasker on Android. No forums, no research, no weekends lost to yak-shaving. Just a need, a conversation, and then the solution—before I even had a name for what I was building.

This keeps happening. Nameless workflows. The space of possible ways-of-working just got much wider than the space of things we can point at and say “that’s a known job.”

Why now?

The bottleneck moved. It used to be execution. “Can I build this?” was the hard question. You had an idea, and the constraint was whether you could actually make it real—or find someone who could, and afford them, and communicate well enough that they’d build what you meant.

Now execution is covered, at least the first 30%. You still have to attend to things. But you’re not writing 80% of the code anymore. Your input is making sure things align with your intent as they go along.

The bottleneck is now exploration. Hunches. Taste. The feeling of being onto something.

Those hunches are becoming self-fulfilling. Not in the hyperstition sense, where you believe something and unwittingly cause it. More like a self-fulfilling lead. You’re onto something, it’s investigative, you’re chasing it down. That very feeling is now a trigger to approach a frontier model.

You describe the workflow you wish existed. The model knows the established options, the common patterns. It gives you the workflow. You ask for something you can hand to another agent. It does that too. You’re left with a list of instructions to set things up on your phone.

You can complete it over a coffee.

People talk about model overhang—capabilities ahead of deployment, things we haven’t figured out how to use yet.

I think there’s a human overhang too. Most people haven’t updated their priors on what’s possible. They’re running 2024 mental models on 2026 capabilities. There are workflows that simply didn’t make sense six months ago which you can set up trivially now.

The space of these nameless workflows is enormous. And almost completely unexplored.

We’re in an odd period where your bespoke workflow works before it gets absorbed into a proprietary harness. The bitter lesson is coming for all this stitching and glue—eventually these compositions will become features of official tooling. But right now? Your hunches are pretty much ready to go if you can describe them to a frontier model.

The question used to be “Can I build this?” Now it’s “Do I have the taste to know what’s worth building?”

Most people are still optimising for the old constraint. Meanwhile, the space of available workflows keeps growing.