Most people use AI to get out of doing work. That’s fine. But there’s a different mode — one where AI makes you better at something rather than just handling it for you — and it’s almost completely ignored in how-to advice.
The distinction matters because these two uses produce opposite outcomes. If you use ChatGPT to write your emails, you’ll always need ChatGPT to write your emails. If you use it to critique your emails, explain what’s weak about them, and then make you rewrite them, you’ll eventually write better emails without it.
Why AI Is Unusually Good at This
Getting feedback on skill development normally requires an expert who’s willing to review your work repeatedly, at low stakes, without getting bored. That’s rare. A good writing teacher, a patient chess coach, an editor who’ll mark up your fifth draft — these are either expensive or socially exhausting to ask for repeatedly.
AI doesn’t get tired of your drafts. It doesn’t have a polite threshold where it stops giving real feedback. Ask it to be harsh and it generally will be, if you’re specific about what you want. Available on demand, infinitely patient, capable of detailed critique — that combination makes it a genuinely useful practice partner for a specific set of skills.
The catch is that it can also just do the work for you, and that’s the easier path, so most people take it.
The Setup That Actually Works
The key is structuring the interaction so the model is in feedback mode, not completion mode. This is a prompting pattern, not a setting — you have to impose it deliberately.
Completion mode: “Write a cover letter for a product manager role at a fintech startup.”
Feedback mode: “I’m going to write a cover letter for a product manager role at a fintech startup. After I paste it, do three things: tell me the single weakest sentence, tell me what’s missing that a hiring manager would notice, and tell me what’s working. Don’t rewrite it — just give me the critique.”
The second version forces the model out of its default helpfulness — which usually means producing output — and into an evaluative role. You still do the work. The AI tells you where the work isn’t good enough yet.
Skills Where This Works Particularly Well
AI feedback is most useful when quality can be evaluated against articulable criteria the model actually knows. Writing, argumentation, code, structured problem-solving, interview prep, language learning. Less useful for skills that are primarily physical, perceptual, or where quality is genuinely context-dependent in ways a model can’t access.
A few concrete setups:
Writing — the Socratic rewrite loop
Write a paragraph. Paste it. Ask the model to identify the single most confused sentence and explain why it’s confused. Don’t accept the rewrite it will probably offer. Fix it yourself. Repeat. The point is building the judgment, not just producing a clean paragraph by the end of the session.
Argument and reasoning — devil’s advocate mode
State a position you hold. Ask the model for its best steelman of the opposing view — not a weak counterargument, but the strongest version someone smart could make. Then ask it to point out gaps in your original argument. This is genuinely good for getting less sloppy about things you think you already understand.
Language learning — production practice with correction
Write something in the language you’re learning, then ask the model to correct it and explain why each correction was made — not just fix it, but give you the rule or the reason. That’s closer to what a good tutor does than Duolingo’s gamified drills. The obvious weakness is pronunciation — you’ll need something else for that.
Interview prep — the uncomfortable version
Give the model a job description and ask it to interview you for the role. Tell it to push back on vague answers and ask follow-ups rather than accepting the first thing you say. Actually answer in text rather than scripting responses in advance. It won’t be as uncomfortable as a real interview, but it’s more useful than rehearsing alone.
Where This Breaks Down
Models can be confidently wrong about domain-specific criteria. If you’re learning a specialized skill — a niche legal area, a specific scientific subfield, a craft with real industry standards — feedback may sound authoritative while missing what actually matters to practitioners. The more specialized the domain, the more you should triangulate against a real expert rather than taking AI critique at face value.
This also requires more willpower than just using AI to do the thing. Every session, you’ll feel the pull to ask it to fix your draft. You have to actively resist that, which sounds obvious but is the main reason people drift back into completion mode. Building a habit where you write first, get feedback second, and revise yourself third takes some deliberate structure — a timer, a rule, something external.
And the model has no memory across sessions unless you’re using something with persistent context. It can’t track whether you’re actually improving on the weaknesses it flagged last time. You have to keep that thread yourself, which makes a notebook or a plain text file more useful than you’d expect.
The Underlying Principle
Skill acquisition requires doing the hard cognitive work, not bypassing it. AI is very good at bypassing it. Using AI for learning means intentionally blocking that bypass — putting the model in a role where it evaluates your work rather than replaces it.
Before your next session, write out what “feedback mode” looks like for the specific skill you’re trying to build. What should the model be evaluating? What should it not be allowed to just do for you? That framing, written down before you open the chat, will do more than any particular prompt trick — and it forces you to be honest about whether you actually want to learn the thing or just have it done.
