How Should You Actually Use AI? A Practical Framework

Most advice about how to use AI comes in two registers. One is breathless: it will 10x your output, automate your job, change everything by Thursday. The other is dismissive: it's fancy autocomplete, a plagiarism engine, a toy. Both are answers to a question nobody should be asking, which is whether AI is good or bad in general.

The useful question is narrower: for this specific task, does AI help, and in what mode? Some tasks you should hand over almost entirely. Some get better through conversation. And some look like time-savers but aren't, because checking the output costs as much as doing the work. Getting this sorting right matters more than any prompting trick.

Here's the framework. Three modes, one question each.

Delegation: when checking beats doing

Delegate a task when you can verify the output faster than you could produce it. That's the whole rule.

Take meeting notes. Turning a rambling transcript into a list of action items takes a person twenty minutes; scanning the AI's list against your memory of the meeting takes two. The gap between those numbers is pure profit. The same goes for first drafts of routine emails, summaries of documents you'd otherwise skim anyway, boilerplate, and translation between formats — a spreadsheet into prose, prose into a table, a policy into a checklist.

What these tasks share: you already know what good looks like, and errors are visible on inspection. If the summary misses the point, you notice in seconds. Hand these over without guilt. Doing them yourself is not diligence; it's nostalgia.

Collaboration: when the value is the back-and-forth

The second mode covers tasks where you don't want a finished product — you want a better version of your own thinking. Weighing a decision. Mapping options you haven't considered. Stress-testing a plan before you present it. Learning a topic well enough to ask smart questions.

Here the right mental model is a colleague: tireless, extremely well-read, occasionally wrong, and — usefully — without ego. You can ask it the same question five ways. You can tell it its answer is weak and demand a better one. You can say "argue the other side" and it will, without sulking. You can admit you don't understand something basic, which is often harder to do with an actual colleague than it should be.

A concrete habit: before any consequential decision, ask for the strongest case against what you're planning to do. Not a list of "considerations" — the actual opposing argument, made with conviction. Half the time you'll dismantle it easily and proceed with more confidence. The other half, you'll find the hole before your boss or your customer does.

The failure mode here is passivity. If you accept the first answer, you're not collaborating; you're delegating a task that shouldn't be delegated. The value lives in the second and third exchange.

Verification-bound tasks: keep your hand on the wheel

The third category is where the breathless advice does real damage. Some tasks are verification-bound: confirming the output takes about as long as producing it, so AI saves far less than it appears to.

Novel factual claims are the classic case. If an AI gives you a statistic, a citation, or what a regulation requires, you cannot check that by rereading the answer — the answer is the thing in question. You have to go to the source, which is what you'd have done anyway.

A contract clause is a good test. AI can summarize the contract, flag unusual terms, explain what an indemnification clause typically does — all genuinely useful. But if the clause matters, you still read the clause. The summary is a map; you don't sign a map. The same logic covers medical dosages, legal deadlines, figures going into a filing — anything where being confidently wrong is expensive.

This doesn't mean avoiding AI here. It means budgeting honestly: in this category, AI accelerates the approach to the answer, not the answer itself.

The trust gradient

Across all three modes, one calibration matters more than any other: match your scrutiny to the stakes and to how verifiable the output is — never to how confident the AI sounds.

Fluency is not accuracy. The polish of the prose tells you nothing about the truth of the content, because the polish is constant.

An AI states its shakiest guess in the same assured tone as its most solid fact. That's not deception; it's how the machinery works. But it means the usual human signal — hesitation as a marker of uncertainty — is missing, and you have to supply the skepticism yourself. So build the gradient deliberately. Low stakes and easy to check — a reformatted table, a draft agenda — skim and ship. High stakes but easy to check — code with a good test suite — let it run, then verify hard. High stakes and hard to check — an unfamiliar legal question — treat the output as a lead, not an answer. The tone never changes; your scrutiny has to.

Prompting: what matters, what's folklore

A few habits genuinely change output quality:

• Give context and constraints. "Write an email" produces mush. "Write a three-sentence email to a client who missed a payment — firm, but keep the relationship" produces something usable.
• Show an example of what good looks like. One sample of your preferred format beats three paragraphs describing it.
• Ask for the reasoning or the counter-case, not just the conclusion. The explanation is where errors become visible.
• Iterate. The first draft is an opening bid. "Shorter, less formal, cut the second point" is where the real work happens.

And then there's folklore: magic phrases, elaborate role-play incantations, the belief that "please" measurably improves answers, hundred-line prompt templates traded like collectibles. Modern models mostly don't need any of it. If your prompt reads like a spell, you're optimizing the wrong thing. Clarity about what you want beats ritual every time.

The skills that appreciate

As models improve, prompting tricks depreciate — each generation compensates better for a vague request. Four things move the other way.

Judgment: knowing which mode a task belongs in, and when the answer in front of you is good enough to ship. Taste: recognizing that a technically correct output is still mediocre. Domain knowledge: you can only verify at the speed of what you know, and expertise is what makes delegation safe. And clear writing — because specifying what you actually want, precisely, turns out to be the durable skill underneath all of "prompt engineering."

None of these are AI skills. They're the same skills that make someone good at directing other people's work. That's not a coincidence.

The point is the division of labor

The goal was never to use AI more. Usage is not a virtue; some of the proudest power users are laundering mediocre judgment through impressive-looking output. The goal is a deliberate division of labor: these tasks I delegate because I can check them cheaply, these I work through in conversation because the friction sharpens my thinking, and these I keep my own hands on because the cost of being wrong is mine to carry.

If you can state that division out loud and defend each line of it, you're using AI well — whatever your usage numbers say. If you can't, no framework, and no model, will save you from the results.