How to use AI for deep work without wrecking your focus

Use AI for Deep Work Without Wrecking Focus

Treat AI as a scheduled collaborator, not an always-on tab. Batch it into deep-work blocks: open Claude Projects with your context loaded, prompt once, think, then close it. Gloria Mark's research shows attention spans have fallen to 47 seconds, so the real risk is fragmentation, not the model.

I keep AI on a leash during my own focus hours, and that one habit has protected my thinking more than any app ever has. The problem is rarely the model; it is the posture. An always-open chat window becomes one more thing pinging for attention, and Gloria Mark's research at UC Irvine found the average attention span on a screen has dropped to about 47 seconds, down from two and a half minutes in 2004, with knowledge workers absorbing roughly 275 interruptions a day. Each switch costs around 23 minutes to fully recover. Bolt a chatty assistant onto that and you do not get deep work; you get a faster treadmill.

Cal Newport defines deep work as cognitively demanding effort performed without distraction, and he argues writing is to thinking what walking is to physical health. The danger with AI is subtle: it offloads exactly the friction that produces understanding. A 2025 study by Michael Gerlich of 666 participants found a clear negative correlation between frequent AI use and critical-thinking scores, mediated by cognitive offloading, and the effect was strongest in the youngest, heaviest users. A Microsoft and Carnegie Mellon survey of 319 knowledge workers across 936 real tasks reached a related conclusion: the more people trusted the AI, the less critical thinking they applied, and the group using AI produced a less diverse set of answers. Convenience quietly narrows the mind.

So I run AI as a scheduled collaborator, not an ambient one. My rule borrows from Kahneman's two systems. Deep work is System 2 territory, slow and effortful, and I refuse to let a tool keep yanking me into fast, reactive System 1 loops. I block ninety minutes, close mail and Slack, and decide in advance the single hard question I am wrestling with. AI enters that block only at defined moments. I will load context into a Claude Project, prompt once, then close the window and think on paper before I read a word back. The model serves the thinking; it does not replace the sitting-with-the-problem that makes thinking happen.

Concretely, three patterns survive contact with real weeks. First, batching: I draft my own raw take first, then ask Claude or ChatGPT to challenge it, never to produce the first version. That order matters, because the Gerlich and Microsoft findings both point to the same failure mode, which is letting the machine think before you do. Second, the adversary move: I prompt the model to argue against my conclusion and list what would have to be true for me to be wrong. That keeps verification, the skill the Microsoft authors say AI is shifting us toward, firmly in my own hands. Third, the closed-loop block: AI is allowed in, then the tab is genuinely closed, not minimized. A minimized assistant is an open loop, and open loops are what shred focus.

The coaching half of my work makes the human mechanism obvious. When I coach a founder who feels scattered, the issue is almost never capability; it is that every tool, including AI, is invited to interrupt at will. We do not add software. We subtract permissions. This mirrors what I have written about the art of discovery: insight arrives in the quiet after you stop consuming, and AI is happy to fill every silence if you let it. The skill is deciding when the silence is the point.

There are honest limits. Some tasks, such as synthesizing forty pages of research, are genuinely better with AI inside the block, and forcing artificial separation there is theatre. The judgment call is whether the task needs your reasoning or merely your supervision. Reasoning tasks earn protected silence; supervisory tasks can stay collaborative throughout. I also watch a private signal: if I cannot explain my own conclusion without rereading the chat, I offloaded too much and the work needs redoing.

The applicable insight is small and durable. Decide, before the block starts, exactly when AI is invited and when it is exiled, then honour that boundary like a meeting on your calendar. Deep work is not the absence of AI; it is the refusal to let AI set the tempo of your attention. Keep the tempo yours, and the tools become extraordinary. Surrender it, and the most capable model in the world will simply help you think shallow thoughts faster.