The art of context switching

Here's what a typical day looks like for me. Between breakfast and bed I check in on about 27 separate pieces of work, run by AI agents, for three different businesses. Seven times out of ten, my next message goes to a different job than my last one. When I'm working, I switch about once a minute.

By every rule of productivity advice, that shouldn't work. It does. In five weeks this way of working finished 141 projects, 118 of them with proof they shipped.

I didn't get better at switching. I redesigned what a switch costs.

• 27

  jobs touched on a typical day, across 6 projects

• 7in 10

  of my messages go to a different job than the last one

• 24

  jobs I steered in one afternoon hour, for all three businesses

• 9words

  how I typically re-enter a job

The research on this is clear. Every switch between tasks costs something, and the cost grows when the new task is complex or unfamiliar.¹ Part of your attention stays stuck on the task you just left, which researchers call attention residue, and it drags down your work on the next one.²

In a well-known study of office workers, people stayed in one project for about 11 minutes before switching. When their work was interrupted, it took them about 25 minutes to get back to it, and they worked on two other things in between.³ And you can only hold about four things in mind at once.⁴

So the usual advice is simple: switch less. Block your calendar, turn off notifications, do one thing at a time.

I went the other way. I made the switch itself nearly free.

I don't keep the state of 27 jobs in my head. The system keeps it, and I re-read it every time I come back.

Every job I pin has its own card on my task board, Helm, with its current state, its next step and a dated timeline the agents keep up to date. When I ask what needs me, my coordinating agent doesn't answer from memory. Its instructions say status reports "go stale within hours," so it rebuilds the picture from every thread, Slack and email, right then.

Each of those cards also has one pinned agent thread of its own, so everything about a job stays in one conversation I can click straight back into. Since I set it up that way on September 22, I finish about 27 cards a week, up from about 11.

Researchers call this cognitive offloading: putting memory into the world so your mind doesn't have to carry it.⁵ A shopping list is offloading. I've done it for three businesses' worth of work.

When I switch from a billing migration to a podcast edit to a sign-up bug, the shape of what I read doesn't change. Every card has the same sections in the same order. Every job on my status page reads the same way, from state to next step, so I take each one in at a glance.

That sameness is the trick. Chess masters can rebuild a real game position from a few seconds' glance with about 93% accuracy. Show them a random arrangement of pieces and they do no better than a beginner.⁶ Their advantage is pattern, not memory: real positions follow patterns they've seen thousands of times. A consistent format lets me read a job the way a master reads a board.

Everything my agents write to me follows one rule: lead with the answer. My AI tools run in a setting that tells them to make the first sentence "what happened" or "what's the answer." The guide my agents follow when they draft my messages to other people sets the same rules:

• "Answer the question they actually asked in the first sentence."
• "Technical detail is opt-in."
• "Ask the smallest unanswered question that moves the conversation forward."

It's not a new idea. The US Army requires the bottom line up front in its writing, with sentences that average about 15 words.⁷ Online, people read at most about 28% of the words on a page,⁸ so whatever comes first is what gets read. And plain writing doesn't make you look less smart. In five experiments, needlessly complex writing made readers rate the author as less intelligent.⁹

A keyword check finds "short," "plain" or "lead with the answer" written into 61 of the 80 skills and rule files I've built. Because the defaults already ask for it, only about 1 in 100 of my messages has to say "make it simpler."

The most expensive question in any job is "what do you need from me?" I almost never ask a single job's agent: it shows up in under 1% of my messages. I ask once, for everything, and the answers are already written down on each job's card and on one page, the Thread Desk, which I describe in the Helm article.

An agent only puts a question there if it truly can't proceed without me and I haven't already answered it somewhere else. I answer right on the ticket, and my coordinating agent passes the answer to the agent that asked. My daily to-do list leaves off anything an agent already owns, so nothing reaches me twice.

Research on notifications points the same way: holding them until a natural break in the task cut people's frustration and sped up their responses.¹⁰ My agents don't ping me mid-thought. They line up at the desk.

When an agent finishes something I can see, it sends me a screenshot. I see it in the chat, and I instantly remember what that job is about. I look at the picture instead of opening the page, and answer.

From September 22 to October 26, agents sent me a median of 46 pictures a day of what they had built, on 34 of 35 days. In four out of five of my busiest conversations, at least one screenshot came up. It goes the other way too: I pasted 438 pictures of my own into my messages, at least one every day.

People are remarkably good at recognizing what they've seen. In one study, after looking at 2,500 objects over five and a half hours, people picked the one they'd seen out of a pair 87% to 92% of the time.¹¹ An earlier study found people remember pictures better than words.¹² A screenshot gives me something to recognize instead of something to recall.

Put the five rules together and a switch looks like this. I come back to a job, glance at the agent's last reply, which leads with the answer, and send one short message. When an agent finishes, I'm typically back within about five minutes.

My first message back is typically 9 words. Most are a single line. About one comeback in seven is just a go-ahead: "yes," "send it," "ship it." The rest steer the work or ask a question.

After a day away, I'm more than twice as likely to open with a get-my-bearings question as when I've been gone under ten minutes. I don't write more. I just ask. That's the whole skill: knowing the one question that gets you back up to speed, and trusting the system to answer it.

In the nine days my dictation app has logged, October 18 to 26, three of every four words I sent Claude were spoken, not typed. A spoken message to an agent is typically 22 words. A typed one is 7. Speaking is about three times faster than typing on a phone,¹³ and I give far more context without trying.

It also means I can work from anywhere. In one week of tracking, I steered 23 jobs in 8 projects away from my Mac, on all seven days, from 7 am to 2 am, while the work kept running on the Mac.

About a quarter of what I type is a quick status check, because asking is cheaper than remembering. Here are real ones:

• "Is it live yet?"
• "What's left before stage 1?"
• "Did he respond? What else do we have to do here?"
• "Is this thread done? And this card?"
• "Rally everything about this and put the situation in front of me."
• "Give me the bird's-eye view. What are some important things that I need to know moving forward?"
• "Catch me up on what we've already talked about for these accounts. Has anyone said anything about this in the Slack?"

And my favorite, sent from my phone on a busy evening: "I'm currently on my phone, and you just gave me a ton to read. Can you just boil it down to core bullet points of what I need to know or what I need to decide on right now?"

That's the habit worth copying. Don't ask the AI to explain everything. Ask it for the one thing you need to know or decide right now.

1. Write each job's state down where you can re-read it in seconds: what it is, what's next, what it's waiting on.
2. Use the same format for every job, so you read them by pattern, not word by word.
3. Tell your AI to lead with the answer. Paste the rules above into its instructions.
4. Put everything that needs you on one page, and stop checking anywhere else.
5. When you come back to a job, ask one short question first. "Where are we?" beats rereading the whole thread.

Context switching isn't a talent for juggling. It's a skill of saying exactly what you need, in as few words as possible, to a system built to answer it. Every rule above comes back to that: write the state down, keep the format the same, put the answer first, and ask one clear question.

That's also where the leverage comes from. In the 206× experiment, I timed every one of those short instructions and had blind AI estimators price the work: one hour of them produced an estimated 206 hours of skilled work. And Helm is the system that holds all of it together.