Most advice on this reduces to "use AI less" or "stay critical". Neither survives contact with a Tuesday afternoon and a deadline. Vigilance is not a plan, because the whole appeal of the tool is that it lets you stop being vigilant.

What follows is a method that does not depend on willpower. It comes out of what the research says about skill atrophy, and out of running the approach with ten knowledge workers for six weeks.

Start from the right unit

The mistake is treating AI as a setting you turn up or down. It is not. The unit that matters is the task, and the same person should want completely different behaviour from the same model across two tasks in the same hour.

Sort each piece of work into one of five modes.

Automate. Mechanical, well defined, you have already proven you can do it. Reformatting, boilerplate, converting between structures, first-pass transcription. Hand it over. There is no skill here to lose and no virtue in doing it slowly. Ask for annotations so you can verify.

Augment. Complex, and in a domain where you are genuinely strong. Here AI should be fast and should argue with you. Skip the basics, go to edge cases, ask it to attack your reasoning. Speed is the point. Your expertise is the guardrail.

Coach. In a domain you want to grow. This is where AI must slow down. It should ask before it answers, and when it does answer it must show the reasoning, not just the result.

Protect. A skill that defines you professionally, that you would be genuinely unhappy to lose. Judgement calls in your field. Writing in your own voice. The analytical move that got you hired. Here AI should refuse to do the work, and instead make you do it and then critique what you produced.

Hands-off. Ethical calls, relationships, taste, anything where the right answer depends on context only you hold. AI can lay out options. It should not choose.

Two things go wrong when people skip this. They apply Coach-mode friction to Automate tasks, hate it, and abandon the whole idea by Wednesday. Or they apply Automate-mode speed to Protect tasks and lose the skill without noticing, because the output still looks fine.

Write your sentence first

The single highest-leverage habit, and the one with the best evidence behind it.

Before you prompt, write one sentence containing your own answer. Not a good sentence. Not a right answer. Just a committed position.

In Bastani's PNAS trial, the group whose tutor withheld answers until they reasoned scored the same on a closed-book exam as students who never used AI at all. The group with the unguarded tutor scored 17% below that control. Same model. The only difference was whether the student produced a thought before seeing one.

Buçinca and colleagues found the same effect in a different setting: forcing people to commit to a judgement before revealing the AI's recommendation cut over-reliance substantially. It also made people slower, and they disliked it. Which tells you something important. The habit that works feels like the habit that is not working.

The mechanism is simple. If you never formed a view, there is no gap between your view and the model's, and it is the gap that teaches. You cannot be surprised by an answer you had no prediction for.

This costs about fifteen seconds. It is the cheapest intervention in this entire article.

Learn to spot the jagged edge

The BCG consultants who used AI were 25% faster on tasks inside the model's competence. On tasks outside it they were 19 percentage points less accurate than consultants with no AI at all, because the model was confidently wrong and they could not tell.

The skill that matters is knowing which side of that edge you are on. And it is maintained by doing the work yourself, which is why heavy AI use quietly erodes exactly the capability you need to use AI safely.

Practical version: before you accept an output, ask what would have to be true for this to be wrong, and whether you would notice. If the honest answer is "I would not notice", you are on the wrong side of the edge and you should not be delegating that task today.

Ask for reasoning, not just answers

When AI does give you an answer in a domain you are learning, demand the working: the steps, why the load-bearing choices were made, and the general principle. Not because it is nicer. Because a bare answer is the worst of both worlds. It interrupts your own reasoning without teaching you anything, so you pay the skill cost and get none of the learning.

A shown-reasoning answer costs you thirty seconds of reading and leaves you with something transferable. This holds even when you are in a hurry. If you must go fast, go fast and get the three-line worked path. Speed and bare answers are different things, and only one of them is worth having.

Notice the silence

Standard AI agrees with you. Nine of ten people in our pilot named hallucination, sycophancy, or generic voice as a top pain point before they had used anything new. One put it plainly:

"They tend to confirm what you are saying instead of looking holistically at an issue and challenging you."

Another:

"Sometimes their answers lack sources or a solid foundation. They also agree with me too much. I'd like more pushback when needed."

If your AI has not disagreed with you in a week, that is data. Either you have mastered everything you asked about, or you have stopped noticing. Ask it directly to find the weakest part of what it just gave you, and to argue against your framing rather than inside it.

Do not let it hand you the critique every time, though. If the AI always supplies the scepticism, you stop generating your own. Every so often, find the flaw first, then ask.

Protect a small number of things, deliberately

Not everything. Three or four skills, chosen because losing them would change who you are professionally.

Write them down. Do them by hand often enough that the muscle stays. Set a rule for each: "AI never drafts this", or "AI critiques, never writes", or "I outline, AI fills, I rewrite".

The specifics matter less than the fact that you decided in advance, when you were calm, rather than at 4pm on a deadline.

Be selective, or you will quit by Wednesday

The mistake that kills this habit is applying it everywhere.

If you make yourself write a hypothesis before asking AI to reformat a spreadsheet, you will resent the whole method within a week and drop it, including the parts that were working. Most of your work does not need any of this. In our pilot, more than half of all real tasks fell into domains where the right move was to go fast and have the AI attack the output afterwards, not to slow down.

Coaching is for the minority of tasks where you are actually learning something. Protect a few skills, coach a few domains, and let the rest run at full speed. Two of our pilot participants finished their work faster than they had before, precisely because the discipline told them where not to think harder.

Doing this without doing it manually

Everything above is a discipline, and disciplines decay. The reason we built TAOS is that nobody sustains per-task mode-sorting by hand for six months.

It rates your expertise per domain, classifies each task into one of the five modes, and varies its behaviour accordingly: fast where you are strong, questions before answers where you are growing, and a refusal to quietly do the work you said you wanted to keep. It runs inside Claude, ChatGPT, Gemini, Cursor and Codex, and the profile follows you between them.

You can connect TAOS to the AI you already use in a couple of minutes. Or read what the pilot actually found, including where it fell short.