Prompting Gets Too Much Credit
I keep seeing people talk about “AI skill” as if it mostly means knowing what to type into the little box. I understand why. Prompting is visible. You can screenshot it, teach it in an hour, turn it into a cheat sheet, and feel like you learned something tangible. But that is exactly what bothers me about the way we are teaching AI right now. The most visible part is getting the most attention, while the harder skills underneath it are the ones that determine whether the output is actually useful. That is the entire point of this iceberg.
Prompting
Prompting matters. Good instructions beat bad instructions, and I use techniques like examples, constraints, iteration, critique, and role-setting constantly. I just think we have made it far more mysterious than it really is. If someone’s primary AI advantage is that they know a 14-step prompt formula, I would not get too comfortable. The models themselves are getting better at understanding what we mean, which means the elaborate ritual around prompting will probably matter less over time.
Learn it. Use it. Just do not confuse it with the whole skill.
Context
This is where things get much more interesting, because AI has a very annoying limitation: it does not automatically know all the stuff you forgot to tell it. Imagine asking AI to recommend a strategy for a company without explaining the company’s goals, customers, competitors, finances, previous decisions, current problems, internal disagreements, or why half the organization still does something strange because of a decision made seven years ago. Then people complain that the answer is generic.
Well...yeah.
Context is the difference between asking, “What should we do?” and saying, “Here is where we are, here is what happened before, here is what we cannot change, here is what success means, here are three things we tried, and by the way, nobody agrees on number two.” That is much closer to how real work actually happens. It is messy, incomplete, political, historical, and full of details that would never appear in the neat little prompt someone shares online. The better the context, the more useful AI becomes. Not because the AI suddenly got smarter, but because you finally gave it enough of the world you are operating in.
Judgment
This might be my favorite layer because judgment is where the relationship with AI gets slightly adversarial.
AI gives you an answer. Fine. Do you believe it? Which part? Why? What is missing? Is the reasoning actually good, or did it just arrive wearing a nice suit? Judgment also includes knowing which AI to use. That is getting harder because there are an absurd number of choices now: different versions and modes of ChatGPT, multiple Claude models, Gemini, Perplexity, Copilot, specialized research tools, image models, coding models, agents, and whatever launched while I was writing this paragraph.
They are not all equally good at the same things, and “I use AI” is becoming about as descriptive as saying “I use software.” A skilled user knows when to change models, change tools, combine them, or stop using AI entirely. That choice matters just as much as the words you type.
Domain Knowledge
This is where I probably disagree most with the idea that AI makes expertise less important. I think the opposite is happening. AI makes it much easier to produce something that sounds like expertise. Give it almost any topic and it can generate a polished explanation with confident language, logical headings, plausible recommendations, and just enough specificity to make you think, “Yeah, that sounds right.” Then someone who actually knows the subject reads it. That person notices the assumption nobody would make in practice. They know the recommendation violates a constraint that never made it into the prompt. They recognize that two concepts the AI casually combined are not actually interchangeable. They see the thing that looks wonderful in theory and would last approximately nine minutes on a factory floor.
I spend a lot of time in manufacturing, and this distinction is painfully obvious there. You can learn the vocabulary quickly. You can even learn enough to sound convincing. But there are things people know because they have spent years seeing machines fail, projects stall, budgets disappear, operators improvise, and perfectly sensible plans run headfirst into physical reality. AI can give you information. Domain knowledge helps you decide what that information means.
Problem Framing
And then we get to the bottom of the iceberg, which I think belongs there because this is where everything can go wrong before AI even enters the conversation. People love to start with the requested task: “Write this report.” “Analyze this data.” “Create a strategy.” “Automate this process.” Those sound like problems, but they are usually activities. The real question is what someone is trying to accomplish and why the current situation is not good enough.
Take the dashboard example. Someone asks AI to design a better dashboard, and maybe AI creates a fantastic one. Beautiful charts, sensible KPIs, clear layout. But what if the actual problem was that nobody trusted the underlying data? Or that managers already had the information but were not acting on it? Or, my personal favorite, nobody knew what decision the dashboard was supposed to help them make in the first place. That is why problem framing sits at the base for me. If you define the wrong problem, the context can be excellent, your domain knowledge can be deep, your judgment can be strong, and your prompt can be brilliant. You can still end up doing the wrong thing extremely well. And that is really what this entire iceberg is trying to say. Prompting deserves a place in AI skill, but it is the part sticking out of the water because it is easy to see. The larger part is underneath: supplying the right context, exercising judgment, knowing enough about the subject to catch what AI misses, and framing the problem correctly before you ask the first question.
That part is harder to teach, harder to copy, and much harder to show in a screenshot. It is also where most of the skill actually lives.
