Skip to content
Sign in
← All notes

The skills that matter most in the age of AI, and how to tell if you have them

Every list of AI-era skills names the same six words and none of them tell you whether you actually possess one. Here is what each looks like in practice, and the test.

Skills8 min read

Judgement. Critical thinking. Adaptability. Communication. Creativity. AI literacy. Every article names these, nobody defines them, and so nobody can tell whether they have one or are simply confident.

A skill you cannot test is not a skill you can develop, because you have no way of knowing whether you improved. So the useful version of this list is not the words. It is what each looks like when somebody has it, and what to check.

Domain judgement: knowing when a plausible answer is wrong

This is the load-bearing one, and it is the least discussed because it cannot be taught in a workshop. It is the ability to look at a well-structured, confident answer in your field and feel that something is off before you can say what.

It exists only where somebody has done the work, made the mistake and seen the consequence. That is why it does not transfer between fields and why it takes years, and it is also why it is now the scarcest thing on this list.

The test: when you last disagreed with a generated answer, could you say specifically what was wrong with it, or only that you did not like it? Specific beats uneasy.

Domain judgement as the ability to sense that a confident, well-structured answer is wrong before being able to explain why. It is built from having done the work and seen the consequence, which is why it does not transfer between fields.
The one skill on every list that cannot be taught in a workshop.

Framing: turning a vague situation into an answerable question

Models reward precision and punish vagueness, which turned question-framing from a soft skill into a direct multiplier on the quality of what you get back. It was always the harder half of analysis; now the gap is visible in the output.

It is also what separates people who find AI transformative from people who find it disappointing. The tool is the same. The question is not.

The test: give the same brief to two people and compare what they asked before starting. The better framer asks about constraints, users and what would count as success. The weaker one asks about format.

The same vague situation turned into two different questions, producing two very different answers. The tool is identical in both cases; the framing is what changed the result.
Why the same tool is transformative for one person and disappointing for another.

Communication under disagreement, which is not the same as writing well

Writing clearly used to be a reasonable proxy for thinking clearly. That proxy is gone, because clear prose is now free and available to anyone.

What survived is the live version: explaining why a conclusion holds to somebody who does not want it to, staying specific under pressure, and changing your position when the other person is right without treating it as a defeat.

The test: in your last real disagreement at work, did you end up with a better answer, the same answer, or a worse relationship? Only the first one counts.

AI literacy: knowing what the tools are reliably bad at

AI literacy is usually taught as what the tools can do, which is the half that markets itself. The valuable half is the opposite: knowing the failure modes well enough to predict them.

That means knowing that models are confident when wrong, that they smooth over ambiguity rather than flagging it, that they are strongest on what resembles their training and weakest exactly where you most need help, and that they cannot tell you what they do not know.

The test: can you name three things you would not use a model for in your own job, and say why? Somebody who cannot has not used one seriously.

What to do with this

None of these improve from reading. They improve from repetitions with feedback, which means the work is to arrange feedback where there currently is none.

  • Write down a prediction before each significant decision, and check it later. This is the only way judgement gets a scorecard.
  • Before generating anything, write the question you are actually trying to answer. Compare it to the prompt you were about to use.
  • Ask for the objection rather than the approval. Approval teaches nothing.
  • Keep a short list of the times a model was confidently wrong in your field. It becomes the most practical AI literacy training available.

Common questions

What are the most important skills for the AI era?
Domain judgement, question framing, communication under disagreement, and knowing what AI is reliably bad at. The common thread is that none can be demonstrated by producing an artefact, which is exactly why they held their value when producing artefacts became free.
How do I know if I actually have these skills?
Test them rather than assume them. For judgement, check whether you can say specifically what is wrong with a bad answer. For framing, compare the questions you ask before starting with those a colleague asks. For AI literacy, try to name three things you would not use a model for in your own work, with reasons.
Can AI skills be learned quickly?
Tool operation, yes, in days, and it keeps getting easier by design. Domain judgement, no, because it is built from consequences over years. That asymmetry is why the tool half is not a durable advantage and the judgement half is.
Are technical skills still worth learning?
Yes, and for a slightly different reason than before. Deep technical knowledge is what lets you evaluate generated output in that domain. Its value as a way to produce things has fallen; its value as a way to catch a confident mistake has risen.

PDP Quest exists because of the problem underneath all of these: when output stops indicating capability, you need another way to know who can actually do the work.

See how verification works →