How to Work With AI: 3 Skills That Matter More Than Prompts

You're building landing-page copy with ChatGPT. You get a slick offer, and in it, a sentence: “Studies show that 67% of customers abandon their cart when they don't see shipping costs upfront.” A number that fits perfectly — specific, credible, ready for a headline. There's just one problem: the model just made it up. Studies and all.

If you run a business or handle marketing, you're probably in one of two places. Either you're already using AI for offers, posts, and sales emails — and every so often you catch the model making things up. Or you're still on the fence, because you've heard about hallucinations and would rather not embarrass yourself in front of a client with a text where half the facts are wrong. Both positions are reasonable. And in both cases, you can deal with AI's limitations — to a large extent. You'll need the three skills described below.

First, though, I need to show you why the problem doesn't sit entirely in the machine. It sits in us too.

Why do we trust machines that make things up?

We were testing our trust in machines long before anyone had heard of ChatGPT. Take this famous study: psychologists Kathleen Mosier and Linda Skitka sat professional pilots down in a flight simulator on a San Francisco to Los Angeles route. Mosier spent her career studying cockpit crew decisions, working with NASA's Ames Research Center. Skitka is a social psychologist who later became known for her research on moral conviction. A trap was waiting for the pilots in the simulator: conflicting signals about an engine fire. The cockpit gauges showed the left engine was on fire. The tablet running the emergency checklist called for shutting down… the right engine.

Three-quarters of the pilots listened to the automation. The cockpit gauges said one thing, their training said one thing — and yet 75% of the crews shut down the healthy engine because the system suggested it. In the control group, working from a paper checklist, only 25% made the same mistake.

The researchers coined a term for this phenomenon: automation bias — when a machine offers a suggestion, people stop searching for and processing information on their own. The automation's cue acts like a shortcut — instead of thinking, we check what the system says and move on. In later experiments, Mosier and Skitka described two kinds of errors: omission errors (the system missed something, so we did too) and commission errors (the system told us to, so we did it — without checking).

When we use AI, we fall into the exact same trap. We skim and miss the made-up data. Or we don't think about it at all — the system hands us a course of action, and we just execute it.

There's good news in this research too. Pilots who felt personally accountable for how they used the automation checked its suggestions more often and made noticeably fewer mistakes. Awareness that you're using an automated system, plus a sense that you have to put your name on the final result — that's where we'll start. But it's only the start. Because there are three more skills for staying in control of that automation.

Skill one: know where the model's knowledge comes from

To get a handle on the model's fabrications, you need to understand how it “talks” in the first place. The mechanism is simpler than it looks.

A language model does one thing: it predicts the next token, meaning a fragment of a word. It gets some text and guesses what should statistically come next. Then it does it again. And again. All of ChatGPT's eloquence comes down to two things:

• The process repeats thousands of times — several times for a single word alone;
• Tens of billions of parameters get factored into calculating that probability — a scale of computation the human brain can't really grasp, which is why it looks to us a bit like magic or intelligence.

But you know what? You do the exact same thing. When you talk, you don't plan a sentence all the way to its last word — your brain feeds you the next words on the fly. And when you tell an old story and your memory is missing a detail, your brain quietly fills it in with something plausible. Can't remember what color the rental car was? It'll insert a made-up color. Same with the waiter's name, or the amount on the bill. Psychologists call this confabulation. Nobody notices, because the filler fits the story. Heck, sometimes we don't even notice it ourselves!

The model confabulates the exact same way. Every single token. But as long as the prediction is grounded in something the model actually “saw” — in its training data or in material you fed it — the output sticks to the facts. When there's a hole in what the model knows, though, the prediction doesn't stop with an error message. It keeps flowing and fills the hole with content that statistically fits: a plausible number, a plausible report title, a plausible expert's name. That's how my “study about 67% of abandoned carts” was born — the model knew what sentences with statistics look like in marketing copy, so it generated one.

Andrej Karpathy, one of OpenAI's founders, puts it plainly: models are dream machines, and hallucinating is their natural state — we just steer the dream with our prompts. The ability to invent a statistic and the ability to write a catchy slogan are the same mechanism. You can't cut out one without killing the other.

Since you can't stamp out the fabrication, what's left is control. And control starts with a question: where does the model get what it's telling you? There are three sources, and each one asks something different of you.

Model weights — what the model shows up with

Weights are the compressed memory of everything the model was trained on — billions of pages of text squeezed into parameters. That's how the model “knows” copywriting rules, LinkedIn post formats, and general knowledge about your industry.

A practical test I recommend to every business owner: ask the model about your own company and your competitors. But ask a model that has no internet access. You'll see, in black and white, what's actually sitting in the weights: sometimes surprisingly a lot, sometimes nothing, and sometimes — the most dangerous case — a mix of truth and confabulation. This one experiment will teach you more about a model's limits than plenty of courses would.

The bad news: you have no influence over the model's weights. When people using AI talk about “training their own AI,” 99.9% of the time they aren't talking about changing the model's weights. That's a process that's expensive, time-consuming, and not worth it for most of us.

Two pieces of good news. First: you can always switch models. Swap ChatGPT for Gemini or Claude. That option gets a lot more appealing once we reach skill number three. Second piece of good news: the model's weights are only one of the things the model draws on when it creates something for you. So let's jump to the other two, which you can actually modify.

Visible context — what you put in yourself

The second source is in your hands: everything you type and paste into the conversation. Questions, instructions, documents, files. This is the model's working memory — and your strongest tool against hallucinations, because a model that can quote doesn't have to guess.

In a marketer's day-to-day, it looks like this. Writing an offer? Paste in the price list, the service description, and two real customer testimonials — instead of assuming the model “knows something” about your business (you already know what it knows: you checked that in the test above). Asking for sales posts? Attach the product page and a list of the most common objections from sales calls. Want an analysis of campaign results? Export the data and hand over the file, instead of describing the numbers from memory.

Feeding context can be tedious. But in everyday work with AI, it's the single most effective known way to limit fabrication — and the shortest path to copy that sounds like your business, not like the average of the internet.

Here again, two pieces of news: good and bad. The bad one is that without practice, we're terrible at supplying context. We assume the model knows something it has no way of knowing, and then we get annoyed when it hallucinates. The good one: prompting frameworks fix this problem (that's skill number three).

Invisible context — what the model pulls in on its own

The third source is the newest and the most treacherous. Modern models can reach outside while they work: searching the internet, reading pages, connecting to your tools through MCP servers (Model Context Protocol — a protocol that lets the model peek into your calendar, your CRM, or your inbox). Everything the model finds lands in the same working memory as your instructions. With one difference: you don't see it happen.

You ask for competitor research before putting together an offer. The model searches the internet and comes back with an analysis. What did it base it on? An industry report — or a forum post from someone with strong opinions and weak sources? If you don't check, you don't know. So build a habit around one control question: “what sources did you base this on — show me the list.” Glancing at a source list takes thirty seconds. Sending a client an analysis built on forum folklore costs a lot more.

Pro tip: tools like ChatGPT have what's called persistent instructions — something the model sticks to in every new conversation with you. Put “always show your sources” into a persistent instruction.

John Lee and Katrina See, researchers of trust in automation, call this stance trust calibration: the skill isn't trusting the machine or distrusting it, it's knowing when it deserves that trust. The three sources give you a simple scale. An answer from the weights: treat it like a well-read friend's opinion — interesting, worth checking. An answer from your own documents: much more reliable. An answer from the internet: only as good as the sources the model found — so check what it found.

And one more trap to close out this section. Research on working with AI shows the explanation paradox: the more fluently and extensively a model justifies its answer, the more we trust it — regardless of whether it's actually right. We treat people the same way: the more confidently someone explains something to us, the more we trust them.

Skill two: jump from chat to files

The second skill requires rethinking what working with AI even is. Most people reduce working with AI to a single mode: the chat window. You type a question, get an answer, copy it into Word. Meanwhile, the tools have split into three levels, and each one is for something different.

• Chat mode is a conversation. The model replies with text, sometimes generates images. You carry the results forward yourself, manually. Great for thinking: brainstorming a campaign, drafting an offer, testing headlines, learning. Limited when the end result needs to be a file, a spreadsheet, or a presentation.
• Task mode works directly on files. The model gets access to a folder, documents, and tools — and instead of describing what could be done, it just does it. Real examples from a marketer's life: it gets a folder of photos and a CSV file, and hands back eighty product descriptions for the store. It gets an export from your email platform, and hands back a report with list segmentation and recommendations. It gets a recording of a sales call, and hands back a note, an offer, and a CRM entry via an MCP server. You stop being the courier between the chat window and the rest of your computer.
• Code mode — the model can't just work on files. It can also write itself a program to help complete the task you're asking for. Let's call this the advanced level for marketing work. Most of us don't need it.

The trouble is, every vendor names these levels differently. Here's a cheat sheet, as of July 2026:

OpenAI

• Chat level: ChatGPT
• Task level: ChatGPT Work (since July 2026). Previously: Codex
• Code level: Codex or ChatGPT Code

Google

• Chat level: Gemini
• Task level: Antigravity 2.0 (since June 2026); before that, Google had no tool at this level
• Code level: Antigravity IDE, Antigravity CLI (since June 2026). Previously Gemini CLI

Anthropic

Anthropic has “always” kept all three modes in one place — an app called Claude. But their layout shifts around a lot too. So:

• Chat level: the Home tab, Chat mode within the chat (before July 2026: the Chat tab)
• Task level: the Home tab, Cowork mode within the chat (previously: the Cowork tab)
• Code level: the Code tab (pretty much always has been)

This list will go out of date faster than I'd like. In June and July 2026 alone, practically every vendor reshuffled its branding. The names will keep changing. The chat–task–code split will stay, though, because it maps onto three genuinely different kinds of work.

Why do I treat the jump from chat to task mode as a separate skill, rather than a technical footnote? For two reasons.

The first is practical: the productivity gap is a step change, not a gradient. In chat, you ask for report text and format it yourself, paragraph by paragraph. In task mode, you get a finished file, and along the way the model looked at the source material in your folder on its own. That's the difference between dictating a letter to someone and delegating your correspondence.

The second reason is psychological. A study by Gerlich (2025, 666 participants) found that frequent AI use is linked to weaker critical thinking, and the mediating mechanism is cognitive offloading — handing mental tasks off to the machine. It sounds like an argument against AI, but the conclusion is more subtle: the problem starts when we hand the machine our thinking. Task mode lets you hand it something else instead — execution. Copying, formatting, pasting, tidying up files doesn't need a brain; it's a tax on thinking. Automation, used well, pays that tax for you, leaving you the energy for the decisions: who we're talking to, what we're promising, what sets us apart. On one condition — you're still the one making those decisions. In task mode, keeping control over what the model pulled into its context matters even more, because it runs longer and more independently.

Skill three: construct prompts, don't collect them

That leaves the skill most guides start with — and I've deliberately saved it for last. A prompt written without understanding where the model's knowledge comes from and which mode it's running in is like a beautifully phrased question asked of the wrong person in the wrong room.

A good prompt has at least three elements.

• Role. When you write “you're an experienced e-commerce strategist,” you're not staging a little play. You're sending the model into a specific region of its knowledge cloud. In the weights, strategist-style text clusters together — the role means the answer gets drawn from that neighborhood, not from the whole internet, where absolutely everyone has an opinion on sales.
• Task. You define the end result: what should come out of it, in what structure, for whom, at what length. Not “write something about our product,” but “write three landing-page headline variants for a prospect who's comparing us to competitors and is worried about implementation costs.”
• Hint. The element people forget most often: something that knocks the model out of writing “like everyone else.” An example of an email that worked for you. A snippet of text in your own style. A counter-example: “avoid a tone where every sentence sounds like a slogan.” Without a hint, the model settles back into the middle of its statistical distribution — the averaged-out language of the internet, the same language your competitors get out of the same model. A hint is your way of saying: not from there, from here.

Role, task, hint are enough for most everyday uses. I show how this logic works in practice across sixteen examples in the article about ChatGPT copywriting prompts.

And when the task is more serious — a communication strategy, an offer for a key client, a campaign plan — I recommend the CRIT structure popularized by Geoff Woods: Context, Role, Interview, Task.

The most interesting element is the third one. In classic prompting, you're the one guessing what the model needs and cramming it all into a single command. CRIT flips the roles: you have the model interview you first. “Before you start, ask me questions — one at a time — until you've gathered everything you need.” The model asks about things you wouldn't have thought of yourself: who makes the decision on the client's side, what the client already tried, why that didn't work. You answer questions instead of writing an essay.

Notice what just happened: we're back to skill one. The interview is filling in visible context — except now the model is running the process and pointing out the gaps in its own knowledge. Instead of confabulating the missing facts, it asks for them. The three skills I'm writing about here aren't separate lessons. They're one system.

Where to start?

If someone asked me how to work with AI, and I had one sentence for an answer, it would be: be the pilot, not the passenger. A passenger boards, trusts, and consumes the result. A pilot knows where the machine's data comes from, chooses the flight mode, and communicates precisely where we're headed. And from Mosier and Skitka's research, we know what sets good automation pilots apart: the feeling that they, not the machine, are the ones putting their name on the result.

For today, three tasks, one from each skill:

• First: ask the model about your company and your competitors — with no internet search. See what's sitting in the weights. Then, the next time you want to use an answer in a client-facing text, follow up: “how do you know that? show me your sources.”
• Second: take a look at task mode. Check out the tools I mentioned, or open the tabs I wrote about. Give an agent one task where it has to produce real files. Product descriptions from photos, for example, the kind I mentioned earlier. Try to get comfortable with this way of working — it's where things are headed.
• Third: take the prompt you use most often and check whether it has a role, a task, and a hint. If the hint is missing — and it usually is — add an example of a text you consider good.

Want to go deeper? Start with the complete guide to artificial intelligence, or jump straight into practice with ready-made copywriting prompts.