Use AI

There’s a point where telling people to “use AI as an assistant” starts to sound a little silly. If I can hand an AI system a piece of work that would’ve occupied me for an afternoon and get back something genuinely useful, I’m going to use that capability. I don’t see any virtue in recreating the same work slowly just so I can say a human did it.

The interesting part comes after the work comes back: can I trust it, what’s it based on, did it solve the problem I actually had or just the one I happened to describe, and if I use the answer, publish it, build from it, automate it or recommend something because of it, what part of that decision still belongs to me? That’s where the complications start.

Let the Damn Thing Work

AI can already do far more than brainstorm titles and tidy a paragraph. It can research a subject, compare competing arguments, draft code, organize a messy project, analyze a page, write usable copy, spot gaps, summarize a pile of documents and help you work through decisions that would otherwise take considerably longer. So let it. There’s no prize for spending two hours doing something by hand if the machine can competently do most of it in ten minutes. Insisting that AI must always remain some kind of junior assistant misses the point of having a capable system in the first place.

What matters is what happens after the labour disappears. AI may be perfectly capable of doing the work, but that doesn’t automatically mean the result should make the decision for you. Suppose you ask it to compare three products and recommend the best one. It comes back with a clean comparison, sensible categories and a confident recommendation, and nothing looks obviously wrong. At that point, I want to know what the recommendation is standing on.

Maybe the battery-life comparison came from manufacturer specifications. Useful, sure, but not the same thing as independent testing. Several websites may repeat the same figure too, which can look like corroboration until you discover they all got it from the same original source. The answer may still be right. I just want to know what kind of “right” we’re dealing with before I lean on it.

The same problem turns up everywhere. An AI-generated website audit can be technically impressive and still misunderstand what the page is supposed to accomplish. A research summary can faithfully summarize weak sources. A piece of code can work perfectly while solving a problem nobody actually had. That keeps pulling me back to one question.

The Wrong Problem Can Still Get an Excellent Answer

This one bothers me more than hallucinations sometimes. A hallucination can be caught. A beautifully executed solution to the wrong problem can waste an enormous amount of time while looking productive the whole way through.

You can ask AI to improve a homepage and get very good work back. It may rewrite the introduction, tighten the navigation, improve the headings and clean up the calls to action. None of that work has to be bad. The annoying part is that the homepage still might not be the real problem. Visitors could be arriving for the wrong reason, the site architecture might be confusing, or maybe the offer itself is weak and you decided the answer was “better copy” before you’d established what was actually going wrong.

AI will quite happily help you improve the wrong thing if that’s the job you give it. So instead of saying:

Improve my homepage.

I’d rather give it the situation:

People are reaching the homepage but not moving deeper into the site. Diagnose what might be causing that before recommending changes.

Now you’re giving the system a chance to test the problem before it starts fixing things.

Give It the Problem, Not Just the Task

A lot of prompt advice focuses on wording. Wording matters. Context matters too. But I think people sometimes spend too much time trying to manufacture the perfect first prompt and not enough time explaining what they’re actually trying to accomplish.

“Write a comparison article” is a task. “Help someone choosing between these three products decide which one suits their situation, and work out what criteria actually matter before writing anything” is a problem. The second version gives the AI something to think about rather than simply something to produce.

The same applies almost everywhere. Give it the constraints that matter. Tell it what’s already been decided and what isn’t open for discussion. If the site architecture is locked, say so. If you want research kept separate from interpretation, make that clear too. The point is to give the model enough of the real situation that it doesn’t have to quietly invent the missing pieces. Even then, it’ll still make assumptions. That’s fine. You just want to be able to see them before they start steering the work somewhere you never intended to go.

Then Let It Push

Once the problem is clear, I don’t see much value in hovering over every tiny piece of the work. Give AI something substantial to do. Let it analyze the material, draft, argue against your position, organize the research, write the code or find the weak spots in a page. Then work with what comes back.

This is where good AI use starts looking less like “prompt engineering” and more like an actual working conversation. A confident answer deserves pressure if the evidence underneath it looks thin. Weak reasoning needs something stronger than another prettier rewrite. And when research comes back suspiciously tidy, I usually want to know what didn’t fit before I start trusting the summary. If the model misunderstood the objective, correct it and keep going.

You don’t have to solve the whole interaction in the first prompt. Quite often the useful part is noticing what needs to be asked next.

Check What Matters Before You Act

There’s no sensible reason to verify everything with the same intensity. The cost of being wrong matters. So does reversibility. A weak paragraph can be rewritten, a broken layout change can usually be undone, and a bad internal-link suggestion is irritating, not catastrophic.

Other decisions deserve more care. The more money, reputation or autonomy involved, the more I want to understand what supports the answer before acting on it. That could mean checking the source behind a factual claim, looking harder at the evidence behind a recommendation, or understanding what an automation will do when something falls outside the normal case.

The point isn’t to keep a human hand on every lever forever. That can become theatre. If a task is low-risk, easy to reverse and a system has shown that it handles it reliably, backing away can be the sensible decision. Human involvement isn’t automatically valuable just because it’s human. Sometimes it improves the outcome. Sometimes it just slows the work down. You have to know which one you’re dealing with.

Where This Shows Up

This way of working turns up all over Output Alchemy, but each subject deserves its own space. With writing, AI can produce polished work very quickly. The harder question is whether the thought underneath the polish is any good. A beautiful rewrite can still strengthen a claim beyond the evidence, remove useful uncertainty or turn somebody’s actual voice into generic sludge.

With websites and SEO, AI can help diagnose, structure and interpret what you’re seeing. I still want to compare that interpretation against the actual site, the search results and real data where it exists. The model can help read the signals. It isn’t the signal. If your next job is actually building the site, that belongs on Build a Website.

And with automation, the question shifts once nobody’s checking every individual action. What’s the system allowed to do? What happens when it gets something wrong? Can the mistake be reversed? When should it stop and hand the situation back? That’s where the practical risk starts to matter.

Use the Damn Thing

None of this is an argument for timid AI use. Quite the opposite. I want AI doing meaningful work because that’s the point. If it can save me time, find something I missed, challenge an assumption I got too comfortable with or take repetitive work off my plate, good. I’m going to use it.

What I don’t want is to confuse a convincing output with a settled question. That part still deserves some attention.

What Still Belongs to You?

AI can do more of the visible work now, and I don’t see much reason to pretend otherwise. I’m happy to hand work over when the system’s good at it. What I’m less interested in handing over without thinking is deciding what problem we’re actually solving, what deserves to be trusted, and what happens after the answer leaves the chat.

I’m fine with the machine doing the work. I just don’t want to sleepwalk through what comes next.