I opened a chat window to check one thing: whether a particular retry library backed off exponentially by default.
Forty minutes later I understood jitter strategies, why thundering herds happen, and roughly how three SDKs handle the problem. I had learned a genuine amount and written no code. The thing I sat down to do was untouched, and I felt fine about it, because the forty minutes had felt like work.
That is the part worth taking seriously. Every other distraction announces itself.
An AI assistant turns into a distraction when you use it to explore rather than to unblock. The fix is not avoiding it. It is naming the specific question before you open the window, putting a limit on how long you are allowed to ask, and writing down the interesting tangents instead of following them. Pointed at a stuck step, an assistant is one of the fastest tools you have. Pointed at a topic, it will happily spend your entire morning, and the whole time you will feel productive.
Social media has a tell. There is a small flinch when you open it during work, a quiet awareness that you stepped off the path. That flinch does more for you than you realise: it is the signal that eventually gets you back.
An AI assistant produces no flinch. You are asking questions, reading explanations, taking notes. Every visible property of the activity matches real work, so the alarm that normally fires after ten minutes of scrolling never fires at all, and you can stay in it until lunch.
It is also better aimed than any feed, since a recommendation algorithm guesses at your interests while an assistant is handed them directly, in your own words. And it never runs out. A subreddit has a bottom. A conversation does not.
Worth naming precisely, since you cannot defend against a failure mode you have not separated out.
You ask one question. The answer is good and contains a term you half recognise, so you ask about that. That answer raises a design question you had not considered. Forty minutes on, you are better informed and have produced nothing.
What makes this hard to catch is that the time was not wasted in the ordinary sense. You learned real things, loosely related to the work. But learning was not the task, and research spiralling passes every check except the one that matters: did the thing you owed someone get closer to existing.
This one is nearly invisible and it costs the most.
You have a document to write. The first paragraph is uncomfortable, so you ask for an outline. It arrives, it is reasonable, and the document now feels begun. It has not begun. You spent ten minutes on a structure you could have sketched in ninety seconds, and the blank first line is unchanged.
Asking for an outline, for options, for help thinking something through: all legitimate moves in some situations, and all perfect procrastination, because they defer the hard part while looking like preparation for it. If you have asked three preparatory questions about a task you have not started, you are stalling with better tooling.
Assistants follow. That is the design. Three messages into a question about database indexes you make an aside about a paper you read, and it will discuss the paper, thoughtfully and at length, without ever mentioning that this is not what you sat down for. Nothing in the tool pulls you back. The whole steering burden sits with you, at exactly the moment your attention is already looking for an exit.
You get an answer that sounds right, you take it, and you move on. Later someone asks why, and you find that you cannot defend it. You have text you cannot reconstruct the reasoning for.
Not strictly a focus problem, but it shows up in the same sessions with the same root: skipping the uncomfortable part rather than getting past a blocked one. The check is simple and slightly unpleasant. Close the window and explain the answer out loud to nobody. If you cannot, you transcribed it rather than learned it.
The default place to use an assistant is a browser tab, which puts it next to every other browser tab. You go to ask one question and you are now in the same window as your email, the forum, and the thing you were not going to look at.
The switch back is where the session dies. You finish the question, look up, and eight other tabs are asking for a click. Any tool that makes you enter a distraction-rich environment in order to use it will leak, however disciplined you are about the tool itself.
This one habit fixes more than any other and takes ten seconds.
Before opening the chat, write the question down somewhere else: notes file, sticky, top of the document. Not a topic, a question, shaped so you will recognise the answer when it arrives. "Something about caching" is a topic and will cost you thirty minutes. "Does this framework cache the response or the rendered output" is a question, and one exchange settles it.
Writing it outside the chat leaves a record of what you went in for, which is the only reliable way to notice drift. When you catch yourself on message nine, look back at the written question. It was usually answered around message two.
Give the asking a short limit before you start. My rule is five minutes for a lookup and fifteen for anything I would call figuring something out. The clock is not sacred, but the limit forces the question to be specific. With twenty minutes of assistant time inside a ninety minute block you write tighter prompts. With unlimited time you meander.
Hitting the limit without an answer is information: usually the question was badly formed, or the thing blocking you is not the thing you are asking about. Both beat another ten messages.
And keep the session timer running while you ask. The asking is part of the focus block, not a suspension of it. Treating it as a break is how twenty minutes becomes an hour.
The tangents are hard because they are often genuinely valuable. You stumble onto something you clearly should understand better, and ignoring it feels like cowardice. So do not ignore it. Park it.
Keep one plain file open all day. When something interesting appears that is not the task, write one line and go back. Ten seconds. Your brain accepts a written note as handled and stops pushing, which an unwritten intention never does.
Read the list at the end of the week. Maybe a quarter still look worth an hour, and those get their own block, where learning is the output and nobody is being cheated. The rest quietly stopped mattering, which tells you what they were: novelty-seeking in a research costume.
Every use of an assistant falls into one of two categories, and they need different rules.
Unblocking is when a specific step is stopping you: the undocumented error, the clause you have rewritten four times, the formula you cannot remember. Use it here without guilt. The step is identified, the answer is checkable, and you go straight back to the task.
Exploring is trying to understand a subject, compare approaches, or find out what you do not know. Valuable, and not a focus session. It has no endpoint, and it should not happen inside a block where you promised yourself a deliverable.
So exploring gets its own block, with its own profile and an honest label, while unblocking is allowed inside working blocks, time-boxed. The failure that started this post was exploring that smuggled itself into a build session by calling itself a quick question.
| | Unblocking | Exploring | |---|---|---| | Trigger | A named step is stopping you | Curiosity or a decision you have not scoped | | Question shape | Specific, has a checkable answer | Open, no clear endpoint | | Typical length | Two to fifteen minutes | Unbounded | | Belongs in | A working block | Its own dedicated block | | Failure mode | Mild, you get back to work | Loses the whole session |
Most of the above works with whatever tool you use. One part resists habit alone: the tab problem, because the environment re-presents every distraction each time you go to ask something.
Deep Focus puts its AI chat inside the app rather than in a browser tab. Small structural difference, real effect: asking a question does not mean entering the place where everything else lives. It takes file attachments (images, documents, audio), and history is saved and synced across devices, which removes the excuse to reopen a browser to find an old thread.
More useful is that access is a property of the Focus Profile rather than a global setting. Three controls, per profile:
It is a Pro feature, and worth being measured about. It does not decide whether your question was worth asking, and it does not stop you exploring elsewhere. It makes the default sane, which is most of what configuration ever achieves: the version of you at 3pm inherits a decision made by the version who was clear-headed.
Some work is damaged by having an assistant available at all.
First drafts. Everything a first draft needs is already in your head or your notes. The assistant is only a way to avoid the sentence. My drafting profile denies access entirely, and it is the single change that improved my writing output most.
Anything you are learning. If the point is to build the skill, being handed the answer defeats the exercise. Retrieval is where learning happens, and asking skips it while feeling like study.
Work you will have to defend. Use the assistant to check reasoning after you have formed it, not to supply it.
Allow it freely for debugging, reference lookups, mechanical transformations, and the moment you have been genuinely stuck.
It can, in a way that is easy to miss, because AI chat is one of the few distractions that feels identical to work. The damage comes from the mode of use: asking open-ended questions inside a session meant for producing something. Used to get past a specific blocked step, it usually protects focus by shortening the stuck time.
Write the question down before you open the assistant, cap the asking at five to fifteen minutes, and keep a separate list where tangents get one line each instead of a follow-up question. The written question is what lets you notice drift, since drift never feels like drift while it is happening.
Block them for drafting, practice and study, and any work where forming the reasoning yourself is the point. Keep them available for debugging and lookups, where being stuck costs more than the interruption. In Deep Focus this is set per profile, so the answer can differ by session type rather than being one global choice.
It gives procrastination better cover. Asking for an outline, for options, or for help thinking something through can be useful, and can equally be a way to defer starting while appearing to prepare. A check that works: if you have asked several preparatory questions about a task you have not begun, the preparation is the avoidance.
Start with a written question, return to it whenever the thread wanders, and end the conversation once it is answered rather than letting it run on. Tools differ in how much help they give here. Deep Focus can prevent off-topic conversations and inject a custom system prompt per profile, so the assistant itself stays anchored to the work.
The uncomfortable part is that those forty minutes were not stolen from me. I chose every message. The tool only made choosing easy and kept the cost invisible until the afternoon was gone.
Decide what you are asking before you ask it. Most of the rest follows.
Get the latest productivity tips and Deep Focus updates delivered to your inbox
