← All Notes

Related: this is the personal, lived-in version of the instruction-design argument in Kingdom: AI — same principle, applied to how I actually run my day instead of how a system gets organized.

Scaffolding is what clinicians call the techniques people with ADHD use to keep executive function working when the internal version of it isn't reliable — calendars, alarms, timers, music, external structure standing in for the regulation the brain won't do on its own. It's also what software engineers call the stuff that turns a raw model into something you can actually trust to run unsupervised — context management, guardrails, checks, the scaffolding that keeps an agent on task instead of wandering off. Same word. Two completely separate fields invented it independently, for the same reason: both of us — me and the model — have the same failure mode. Scattered attention. Losing the thread. Locking onto whatever's directly in front of us instead of the actual task. Models act a bit like people with ADHD. I've been running the same prescription on both of us for a while now, and it works.

What Scaffolding Actually Grounds

Here's the part that surprised me. I expected AI to help with the obvious stuff — task lists, reminders, keeping a project on track. It does that. What I didn't expect is that it can also do something closer to what people chase with recreational drugs like marijuana: it can ground the more fluid, scattered states of consciousness instead of feeding them. A wandering brain handed something to actually engage with — a real problem, a real question, back-and-forth that requires the analytical part of the brain to stay switched on — comes back down to earth instead of spinning further out. It's not therapy. It's just something to push against that pushes back, on demand, at 2am, with no patience limit and nowhere else to be.

Five Tabs, Cycling

In practice, this looks like a command shell with multiple Claude Code sessions open, each one an agent working its own thread, me cycling between them giving the next instruction, checking what came back, redirecting. Three to five running projects at a time is the sweet spot — enough parallel threads that the ADHD brain has somewhere to go when it wants to move, not so many that nothing gets finished. Seven is my real ceiling, the point where I can still hold the state of every tab in my head without losing one. Ten would be insane — full stop, don't do it.

I'll admit the number in front of me right now, honestly, because this piece is supposed to be honest: nine tabs open in my shell as I write this. Not all of them running agents. Most of them are.

Why This Shouldn't Work — And Does

Here's the tension, and I want to sit in it instead of skating past it: the research on task-switching is not kind to what I just described. The American Psychological Association's own summary of multitasking research puts productivity loss from context-switching as high as 40 percent — every switch has a cost, the brain has to reload the rules of the new task, and that reload isn't free. By that math, nine open threads should be a disaster. It isn't. So either the research is wrong, or what I'm doing isn't actually the thing the research measured.

It's the second one. The switching cost comes from having to reconstruct a task's state in your own head every time you land back on it — what was I doing, where did I leave off, what was the plan. That reconstruction is the expensive part, not the switch itself. Scaffolding is specifically the thing that eliminates the reconstruction: every agent already holds its own context, its own progress, its own next step, sitting there exactly as I left it. Tabbing back in isn't reloading a cold task. It's re-attaching to a thread that never went cold in the first place. The research measured people switching between tasks with nothing holding their place. I built something that holds the place for me. That's the entire difference, and it's the entire point of scaffolding — clinical or technical, human or model.

A dark terminal window showing several glowing session tabs arranged in a row, each one a distinct thread of light, a hand poised over the keyboard about to switch between them

API-Everything

There's a second layer to the scaffolding that took me longer to notice, and it's a genuine shift in how I work: if I catch a CLI agent doing something routine — the kind of task that's going to come up again — my old instinct was to have it write a script. I'm mostly over that now. A script is a one-off crutch. It runs once, on my machine, in my terminal, and does nothing for anyone or anything else. An API is scaffolding that outlives the moment it was built for — callable, reusable, sitting there for the next agent, the next project, the next version of me to lean on without rebuilding it. Why write a simple script when you can build an actual endpoint? APIs are my dope now. Every routine task is a candidate for infrastructure, not a one-time fix.

What to Do Right Now

If you've got the scattered-attention brain — clinically ADHD or just wired that way under deadline pressure — stop trying to fix the wandering with willpower, and stop assuming multiple threads means you're failing to focus. Build the scaffolding instead. Track state externally so a switch is a re-attach, not a reload. Bound each thread so it has a real stopping point instead of running forever. And when something becomes routine, don't script it once and forget it — build the thing that keeps working after you've moved on. That's the whole prescription. It doesn't require fixing the brain. It requires building around it.

1:1 With JP

Running five agent threads without losing your place isn't a personality trait — it's a system, and it's teachable. $100/hr on Zoom, $250 for three hours. On-site in NYC, Jersey City, or Hoboken adds a flat $50 travel fee.

→ Book a session

— J.P. Howlett

Related:

Sources