Most people who say AI gives them mediocre results are right. The results are mediocre. The part they get wrong is the cause. I have spent the past year directing AI to build my website, my team dashboard, my shared memory and a handful of free tools, and the one thing I am sure of is this: the output is a mirror. Here is what I watched, what I got wrong myself, and what I do now.

The day I watched how people use AI

A client, and separately a friend, once showed me how they use Claude and ChatGPT. I watched closely, because I wanted to see what the experience looked like from the other side of the screen.

They were using them as chatbots. They typed a question, waited, and expected a good result to come back. Nothing more went in: no goal, no reader, no context about the business, and no second round when the answer came back average.

I did not say anything in the moment. It was not a criticism of them, because most people were taught to use a search box exactly this way. But it stayed with me, because it was not really about one chat window. It was about what we think we are doing when we use AI.

When I use the same tools, I do something different. I start a conversation. Who is this for? What do they already believe? What should they do after the last line? What does the brand sound like? What should we leave out? Then I let the AI draft, read it, say what is wrong, and let it draft again. The tool is the same. The direction is not.

Three things people believe about AI, and why I think they are wrong

I have heard the same three beliefs enough times now that I can tell which one a person is about to say.

"I already know how to use ChatGPT." Many people do know how to open it, type into it and read the answer. That is not the same as knowing how a large language model works. It predicts likely text from the context you give it. Give it little context and it gives you the most average answer it can find, delivered confidently. It can also be wrong, and it will not always tell you. If you do not know that, you are using a powerful tool the way you would use a vending machine.

"AI does the work, so it is not creative." This is the one I disagree with most. When I build a website with AI, the AI types the code. I decide what the site is for, who it is for, what it should feel like, what to leave out, which version is good enough and which one only looks good enough. The typing was never the hard part. The decisions were. Using your ideas to bring what is in your head into the world, with a very fast assistant doing the execution, is creative work. It is just creative work that nobody sees you doing.

"A longer prompt is a better prompt." Length is not direction. I have seen enormous prompts produce generic output and a three-line message, sent in the right order after the right context, produce something sharp. What helps is a clear goal, real context, examples of what good looks like, and a willingness to go another round. Anthropic's own prompting guidance says much the same: be clear and direct, give context, show examples and iterate.

Notice what these three beliefs share. Each one treats AI as something that should hand you a finished result with no effort from you. That expectation is where mediocre output is born.

Why the output is a mirror

If you give an AI a vague request, you get a vague answer. If you give it a sharp point of view, you get something that carries that point of view. It reflects you. That is why two people can use the same tool, on the same day, and walk away with completely different results.

This is the uncomfortable part, and I want to say it gently, because I have been on the wrong side of it. AI does not take creativity away from people. It removes the place where a lack of direction used to hide. When writing, designing and building were slow, you could look busy while producing nothing original. Now the first draft is instant, and the only thing left to judge is whether you knew what you wanted.

AI does not take creativity away. It removes the place where a lack of direction used to hide.

What directing AI actually looks like

Here is what I have built with AI, and where my part was.

nomadnotepad.app is a notepad that lives in your browser or on your desktop and builds a knowledge graph from your notes, so you can ask your notepad what you wrote years ago. It is not live yet. I will release it when the end-to-end testing is finished, not before. That sentence is also part of the point of this article.

My own numbers, so you can calibrate against them. On a typical page or tool, I accept the result after about three to five rounds of revision. Most of my time on a build goes to directing and checking, not waiting for the AI to type. And on some builds, though not all, I catch something wrong before it ships. These are my own estimates, not tracked counts.

The day my tool said it had submitted everything

I built a helper I call Citations Builder. It is an automated tool that submits a business to directory and listing sites such as DesignRush and GoodFirms. If you work in SEO, you know why: consistent listings build the kind of footprint that helps a business get found, and doing them by hand is slow.

The AI built it. I ran a batch of submissions, and the tool reported success. It said the listings were submitted.

They were not.

Every one of them was sitting in the submission window with no progress, and it stayed that way for days. A local worker on my desktop was already deployed and running automatically, so by every sign I could see, the work should have been moving. It was not.

It looked fine. The output was clean, the status messages were cheerful, and everything on the screen said the work was done. If I had trusted that screen, I would have moved on believing a batch of listings existed that did not. I only found out because I went and looked at where the listings actually were, instead of at what the tool said about them.

I caught it in time. Nothing was published that should not have been, and no client was told something false. But I want you to sit with how close it was, because the failure had nothing to do with the AI being bad at code. It did what these tools do. It produced a confident, plausible report. The mistake would have been mine, for believing the report without checking the world.

This is the same lesson I learned years ago in link building, in a different costume. In my story about shipping fast and getting it wrong, I trusted a number instead of looking at the thing behind it. With AI, the number is the AI's own summary of what it did. Both times, the fix was the same: go and look.

What I check now before I trust AI output

I did not become careful by temperament. I became careful by writing it down, because I have learned that a written standard beats good intentions.

1. Write the standard before I prompt. What does "good" look like here? What would make me reject it? If I cannot answer that, I am not ready to ask.
2. Give real context. Who it is for, what has already been decided, what to avoid. This is why I keep a shared memory across my assistants, which I wrote about in what happens when all your AIs share one memory.
3. Ask it to show evidence. Not "is it done?" but "show me what you did and how I can verify it." A claim without evidence is a guess.
4. Test the real flow. Run it the way a customer would, including the awkward cases, and look at where the work actually ended up. Reading the code or the answer is not testing.
5. Verify against the original. Facts, figures, links and sources get checked against the source. If I cannot find the source, the claim goes.
6. Go another round. The first output is a draft. I treat it like one.

None of this is advanced. It is the same discipline you would apply to a very fast, very confident new hire whom you have not yet learned to trust.

What I mean when I say AI exposes people

I am not saying this to be harsh. I am saying it because I think it is useful to hear.

If your AI results feel mediocre, the answer is not necessarily that you are not creative. It may simply be that nobody ever made you write down what you want, make a decision and defend it. Most jobs did not require it. AI does. That is not an insult. It is an invitation.

The people I see getting the most out of AI are not the ones with the cleverest prompts. They are the ones who know what they are trying to make, who can tell good from almost good, and who are willing to say "no, again." Those are old skills. Taste, standards, curiosity and ownership have always been the scarce part. AI just removed everything else.

If that sounds like a lot of effort, it is. It is also exactly what makes the work yours.

What this means when you hire a virtual assistant

This is where it connects to the company I run.

Many founders now ask me whether they still need a virtual assistant when AI can do so much. I think they have the question backwards. AI is a fast pair of hands. A good VA is a person who directs those hands, checks the result and stands behind it. The skill I look for is not "can use ChatGPT." It is whether they:

• Ask good questions before they start, so they know what the work is for.
• Check before they send, instead of forwarding whatever came back.
• Iterate in rounds, instead of accepting the first draft.
• Own the result, instead of blaming the tool.

A VA with those habits, using AI well, can deliver more, faster and with fewer errors. A VA without them, using AI, produces mistakes at higher speed. That is the difference we try to screen for in how we vet, and it is why a clear brief matters so much. If you want to see what a good brief looks like, I wrote it up in how to brief a content writer, and the same thinking applies when you are briefing an AI. If you are weighing a human writer against AI for your own content, the research and test data are in content writer vs AI in 2026.

For founders who want delegated work done by someone who will direct AI and check it, a content writer, an SEO VA or an executive VA with those habits is worth far more than the cheapest hire who treats AI as a vending machine. Nomad VAs is where I would start for any of the three.

Where to start if you want better results from AI

You do not need a technical background to improve. Here is the path I would follow:

1. Stop typing like it is Google. Describe the goal, the reader and the context before you ask for anything.
2. Write one paragraph on what "good" looks like before you start, and use it to judge the first draft.
3. Expect three to five rounds, not one. Tell it what is wrong in plain words.
4. Make it show its work. Ask how you can check what it claims.
5. Test the real thing, not the summary of the real thing.
6. Keep the decisions yours. What to make, who it is for and what to leave out are the part only you can do.

AI does not replace creativity. It just stops letting anyone pretend.

I will keep writing about what I build and what breaks. If you want a human assistant who works from a clear brief and checks before they send, book a discovery call.

FAQ