Is deep work with AI the same as using AI to do the work for you?

No. AI-assisted deep work means using AI before the session to prepare conditions for concentration—loading context, clearing open loops, defining the exit point—and then doing the cognitive work yourself during the session. The AI handles the preparation phase; the depth of thinking is yours.

How much AI use is too much before a deep work session?

If your pre-session runway takes longer than ten minutes, something is wrong—either your inputs are too scattered, or you are using preparation as a substitute for starting. The runway should take five to eight minutes. If it regularly exceeds that, examine whether you are using it as a delay tactic.

What if I cannot afford a 90-minute block? Can shorter sessions still be deep work?

Yes, though with diminishing returns. Cal Newport advocates for sessions of at least 90 minutes for the kind of output that compounds over time, but 60 minutes can produce real depth if the entry is efficient. Below 45 minutes, the runway cost becomes a significant fraction of the session, and it is harder to enter a genuine flow state. Protect the longest blocks you can, and use the runway to make them count.

Is there research supporting AI-assisted deep work preparation?

The research support is for the underlying mechanisms, not AI specifically. Sophie Leroy's work on attention residue supports interruption triage. Cowan's working-memory research supports context loading. Csikszentmihalyi's flow research supports exit-point definition. AI speeds up these mechanisms; it does not invent them.

Deep Work with AI Assistance: Every Question Answered

The Basics

What exactly is deep work with AI assistance?

Deep work, as Newport defines it, is cognitively demanding work performed in a state of distraction-free concentration. AI assistance means using AI tools in the preparation phase—before the session—to reduce the entry cost. Specifically: loading working-memory context via AI briefing, triaging interruptions via AI-generated resolution strategies, and defining a concrete exit point. During the session itself, AI is closed.

It is not AI-as-collaborator. The work is yours. The runway is AI’s job.

Why does deep work need a runway at all?

Because entering a state of deep concentration is not automatic. The standard advice—“protect the time, block the calendar”—solves the scheduling problem. It does not solve the entry problem.

Research on attention residue (Leroy, 2009) shows that cognitive threads from previous tasks persist into the current one, degrading performance. Research on working memory (Cowan, 2001) shows that complex work requires loading the relevant context into a constrained set of memory slots before progress is possible. Without explicit preparation, both of these entry costs are paid during the session—at the expense of depth.

How is this different from just doing deep work normally?

Deep work done well already includes informal versions of these three preparation elements. What the runway makes explicit is what good practitioners do implicitly: they review their notes before starting, they resolve pending items, they know what they are building.

AI makes those implicit practices explicit, faster, and more thorough. The runway takes the same preparation that a disciplined knowledge worker does naturally and compresses it from twenty to thirty minutes of informal preparation into five to eight minutes of structured AI-assisted preparation.

Setup and Implementation

What AI tool should I use for the runway?

Any competent language model works: Claude, GPT-4, Gemini. The choice matters less than the quality of your prompts and the consistency of the process. Use whatever you already have access to and will actually use.

How do I start if my notes are a mess?

Start with what you have. Paste whatever context exists—even a rough task description and a few bullet points—and let AI work with imperfect inputs. The Gate 1 brief will be less sharp than if you had clean, organized notes, but it will still improve on cold-starting with nothing.

Over time, leave cleaner handoff notes at session end. That is the real solution to messy inputs: better outputs from the current session become better inputs for the next one.

How long should the runway take?

Five to eight minutes for someone who has done it a few times. First week: expect eight to twelve minutes. If it is taking twenty minutes, you are over-elaborating the context or using the runway as a delay. Tighten it.

Can I run the runway the night before instead of right before the session?

Gate 3 (exit point definition) can be done the night before—in fact, pre-defining your exit points the evening before is a useful habit that pairs well with a shutdown routine.

Gate 1 (context loading) is best done immediately before the session. The working-memory brief is most effective when it is read and then immediately acted upon. If you load context the night before, you will have to reload it in the morning anyway.

Gate 2 (interruption triage) should be done immediately before, because the open loops that need resolution are the ones that currently exist—not the ones you anticipated the night before.

What if I do not have ninety minutes? Can I still use the runway?

Yes. The runway is not calibrated to session length; it is calibrated to entry conditions. If you have forty-five minutes, run the same three gates but set a tighter exit point in Gate 3. The runway will take the same five to eight minutes regardless of session length.

During and After the Session

Should AI be open during the session?

No. Close it before you start. Not minimize—close.

If AI is visible, you will use it. Using it is a context switch. Context switches generate attention residue. Attention residue degrades the depth you just spent five minutes building.

What if I need to look something up during the session?

Two options. Option 1: pre-load the relevant reference information during Gate 1 so it is in your notes before you start. Option 2: write the question on a physical notepad and resolve it at the session break, not mid-session.

For research-heavy technical work, a rule of “lookup only, maximum two queries, immediately closed after” can work—but it requires discipline, and most people find it slips into something more conversational. The cleaner rule is no AI during the session.

What if I get interrupted anyway?

If an interruption occurs, use the mid-session recovery prompt (Prompt 4 in our AI prompts guide): describe what you were working on before the interruption and ask for a three-sentence re-orientation. This collapses the typical 15-minute recovery into two minutes.

Then close the AI again and resume.

How do I end a session?

Write your handoff note before you close the document. Three sentences: what you completed, where you left off exactly, first action next session. This is the input for the next session’s Gate 1 prompt.

Do not use AI to summarize your session before you have written your own notes. Write your own synthesis first. AI can then help structure it—but your unmediated reflection should come first.

Common Problems

The runway feels like procrastination. How do I know the difference?

Runway: five to eight minutes of structured prompting that leaves you oriented and ready to start. Procrastination: twenty minutes of AI interaction that produces a lot of output but no sense of where to begin.

The clearest test: when you finish the runway, do you know the first sentence you will write or the first function you will build? If yes, it was a runway. If you are still uncertain what to start on, you were delaying.

My sessions feel productive but my output is thin. What is happening?

Almost certainly: AI is too active during the session. If you are iterating with AI while working, you are producing AI-mediated output rather than output from sustained solo concentration. Reduce mid-session AI use to zero for two weeks and compare the output quality.

I always hit Gate 1 (context loading) but skip Gates 2 and 3. Is that okay?

It is better than nothing, but you are leaving value on the table. Gate 2 handles the anxiety that quietly degrades sessions even when you believe you have mentally set it aside. Gate 3 handles the goal ambiguity that causes sessions to drift.

Gate 1 alone reduces ramp time. All three gates together change the character of the session.

I lose focus after about 30 minutes even when the runway is complete. What is wrong?

A few possibilities. First, your challenge-skill ratio may be off: the task is either too easy (boring, not engaging) or too difficult (anxiety-inducing, triggering avoidance). Second, your session may be scheduled at a non-optimal time for your cognitive peak—see the deep work scheduling guide. Third, you may have a chronic sleep deficit or physical state issue that prevents sustained concentration regardless of preparation.

Preparation solves entry. Sustained concentration requires adequate cognitive resources and appropriate task difficulty.

The Research Behind It

Is there scientific support for the runway approach?

The support is for the underlying mechanisms, not the specific AI implementation. Sophie Leroy’s attention residue research (2009) directly supports interruption triage. Cowan’s working-memory research (2001) supports context loading. Csikszentmihalyi’s flow research (1990) supports exit-point definition. Ericsson’s deliberate practice research supports the principle of closed-loop, focused sessions over partially attended work.

For a full treatment of the research, see our science of deep work assistance article.

Is there research specifically on AI-assisted deep work preparation?

Not yet—the literature on AI-augmented knowledge work is nascent, and most published studies focus on AI’s effect on output quantity rather than depth of cognitive engagement. The runway framework is built on well-established cognitive science with AI as the implementation tool.

This is worth flagging honestly: we are applying established research about cognitive preparation to a new tool. The mechanism is solid; the specific AI implementation has not been studied in controlled settings.

Does using AI for context loading weaken my ability to do it manually?

An open question without definitive evidence. The concern is reasonable—if AI consistently handles context reconstruction, the skill of building robust mental models through deliberate re-engagement might atrophy.

The conservative position: use AI for context loading when it saves meaningful time (returning to work after 24+ hours away), not as a replacement for the daily habit of leaving clear session notes for yourself. The handoff note habit, maintained independently of AI, ensures you are building the underlying skill.

Getting Started

What should I do today if I want to try this?

Before your next scheduled focus session, run only Gate 3. Paste your task description into AI and ask it to define a specific, observable exit point. Write that exit point somewhere visible. Start working.

That single prompt will give you a clear before/after comparison. Most people find the session noticeably more directed. From there, add Gate 1 next session, then Gate 2 the session after. Build the runway incrementally.

Related:

Tags: deep work, FAQ, AI assistance, focus, knowledge work

Frequently Asked Questions

Is deep work with AI the same as using AI to do the work for you?

No. AI-assisted deep work means using AI before the session to prepare conditions for concentration—loading context, clearing open loops, defining the exit point—and then doing the cognitive work yourself during the session. The AI handles the preparation phase; the depth of thinking is yours.
How much AI use is too much before a deep work session?

If your pre-session runway takes longer than ten minutes, something is wrong—either your inputs are too scattered, or you are using preparation as a substitute for starting. The runway should take five to eight minutes. If it regularly exceeds that, examine whether you are using it as a delay tactic.
What if I cannot afford a 90-minute block? Can shorter sessions still be deep work?

Yes, though with diminishing returns. Cal Newport advocates for sessions of at least 90 minutes for the kind of output that compounds over time, but 60 minutes can produce real depth if the entry is efficient. Below 45 minutes, the runway cost becomes a significant fraction of the session, and it is harder to enter a genuine flow state. Protect the longest blocks you can, and use the runway to make them count.
Is there research supporting AI-assisted deep work preparation?

The research support is for the underlying mechanisms, not AI specifically. Sophie Leroy's work on attention residue supports interruption triage. Cowan's working-memory research supports context loading. Csikszentmihalyi's flow research supports exit-point definition. AI speeds up these mechanisms; it does not invent them.