One of the things you least expect when working with AI is how reluctant it can be to follow your instructions.
Each model has something like a personality. Some are eager. Some are literal. Some are better at following directions. As the models improve, they generally get better at this, but there is still something strange about telling a computer what to do and then watching it not quite do it.
You expect a machine to obey because it is a machine.
But the more we give models the ability to think freely — or at least to simulate thought — the less they behave like the old deterministic computers we grew up with.
This brings us back to an old human problem: how do you get someone to do what you actually want them to do?
With people, you have psychology, incentives, identity, emotion, prior commitments, and trust. With models, you do not have all of that in the same way. But both humans and models share one requirement.
You have to be clear.
Models will hallucinate, but they will also do something more mundane: they will claim they completed a task when they did not. They will hand-wave. They will do half the job, stop, and tell you the whole thing is finished.
So you have to ask for proof.
You need the system to show what changed, cite where it found the evidence, run the check, produce the diff, or demonstrate that the instruction was actually followed. Without proof, the model can drift into the easiest story: “Done.”
It is a strange thing to wrestle in language with a computer.
I recently started teaching my children to code. To do that, I created a custom Claude agent informed by pedagogical research. One of the first things it teaches is that when a program has a bug, the bug usually comes from improper instruction. The machine follows the sequence it was given. If the result is wrong, look at the instructions.
That was the old paradigm of computing: rigid, explicit, unforgiving instructions.
Now we communicate with computers through natural language. The exact thing code tried to remove — ambiguity — is back in the picture.
And so we have powerful new resources, but also new arbitrariness. The future of work is not only about technical skill on one side and human skill on the other. That barbell framing is too simple.
In the age of AI, soft skills and hard skills converge.
Communication becomes operational.
Clarity becomes technical.
Proof becomes infrastructure.