Say What You Mean
Prompting did not die. It fragmented into four different skills. Most of us are still practicing only the first one.
We dismissed it. Most of us did. We heard the word “prompting” and filed it somewhere between data entry and password resets: necessary, perhaps, but certainly not a leadership skill. Not something worth our afternoon. And for a while, we were right. Prompting really was a simple thing. We typed a question, got an answer, refined until the answer was useful. That was the whole of it.
Then the whole of it quietly became several different things, and we did not notice because the word never changed.
IEEE Spectrum declared prompt engineering dead. They were half right. The thing did not die so much as split into pieces, and the pieces turned out to be more interesting than the original.
The first piece is conversation: the ability to hold a useful back-and-forth with an AI system. This is the one most people practice. It is genuinely valuable, and it is also the floor, not the ceiling.
The second is intent engineering: expressing a goal so clearly that the system produces what we actually meant, not merely what we literally said. Andrej Karpathy drew a useful distinction here: prompt engineering tells the model what to do now; context engineering tells it what it knows; intent engineering tells it what to optimize for over time.
The third is system specification: writing the governing documents that tell autonomous agents how to behave over hours or days of independent work. Not a single prompt. A constitution. Operating principles that hold up when nobody is watching, because nobody can watch.
The fourth is meta-prompting: designing how agents prompt each other. When one AI system delegates work to another, someone had to write the protocol for that delegation. That someone is increasingly a human who understands intent at every layer.
Four genuinely different capabilities, all answering to the same name. The field moved on. The vocabulary did not.
The distance between someone who grasps this and someone who does not is already measured in orders of magnitude. Not percentages. Orders of magnitude.
The Skill We Already Had
Here is a paradox that should comfort every experienced leader: the most important of these fragments, intent engineering, is something we have been practicing for years. We just practiced it on people. The skill finally has a name, and the name sounds technical enough to intimidate the very people who are best at it.
Every time we wrote a project brief clear enough that three different teams could execute against it without contradicting each other, we were engineering intent. Every time we delegated with enough context that our people could exercise judgment without checking back on every small decision, we were doing precisely what the new discipline requires.
In the previous articles we covered specification, judgment, decomposition, and orchestration. Each turned out to be an old skill in new clothes. Intent is the same, only more so: it is the skill of communicating what we actually want, completely enough that a system produces it without further guidance. Not a longer prompt. Not more words. Clearer words. The difference between a recipe that says “season to taste” and one that says “half a teaspoon of salt”: both technically correct, but only one works when the cook has never tasted the dish before.
The irony is almost too neat. The people best equipped for this work are often the ones who assume they have nothing to learn. We hear “prompting” and dismiss it as beneath our strategic altitude. By the time we recognize that intent engineering is leadership wearing unfamiliar clothes, the people who started earlier have built months of compounding advantage. The K-shape, once again, rewards the curious over the credentialed. The experts arrive late because expertise told them not to bother.
What Changed
The skill is not new. The medium is. And the medium is unforgivingly literal.
In human communication, the gap between what we say and what we mean is filled by shared context, tone, body language: the whole apparatus of social intelligence we deploy without thinking. Our teams read between the lines. We have spent entire careers relying on ambiguity to do our work for us, and we called that efficiency.
AI has none of that apparatus. It has our words. Everything we leave unsaid, it fills with its own best guess. Sometimes brilliantly. Sometimes an agent sends hundreds of unwanted messages because someone configured a workflow for “outreach” without specifying what outreach did not mean. Same technology. Same generation of model. The difference was the width of a single well-written sentence.
The whole point of human talent now is to simplify. Not to write longer instructions. Not to pile on guardrails after the fact. To say what we mean in the fewest words that cannot be misunderstood. Forty words that leave no room for misinterpretation will outperform four hundred that bury the goal in a hedge of caveats. Brevity is not the enemy of rigor. It is the proof of it.
From Instructions to Governance
As agents grow more autonomous (and they are: current-generation systems work for hours against specifications, unsupervised, at three in the morning), the skill of communicating intent stops resembling prompting and starts resembling governance. The better we get at instructing machines, the more we find ourselves doing something that looks suspiciously like leading people.
Instructions are sequential: do this, then that. Policy is principled: here is what matters, here is what is off limits, here is how to exercise judgment when the instructions run out. The shift from instruction to policy is the same shift every leader makes when a team grows past the point where we can personally review every output. Orchestration (the previous skill in this series) tells us how to coordinate the work. Intent tells us how to govern the decisions.
In Montreal, where Mila and Yoshua Bengio’s research shaped much of the foundation these models stand on, precision of expression carries a particular resonance. Growing up bilingual teaches something monolingual cultures tend to underestimate: saying the same thing in two languages forces us to know what we actually mean. Translation is not repetition; it is discovery. The leap from that discipline to intent engineering is shorter than it looks. It is a cultural inheritance finding a new application.
The hard thinking was always necessary. We were always making these decisions about scope, boundaries, and judgment calls. We just made them implicitly, one conversation at a time, across weeks of back-and-forth. Intent engineering does not add cognitive work. It moves the work forward in time. We think carefully once, specify clearly, and the system executes at a speed that would have seemed absurd two years ago.
And the skill compounds. The first few specifications take longer than expected. The tenth takes half the time. By the twentieth, we have a vocabulary for intent that makes each subsequent project faster: institutional clarity, built through practice rather than training.
The Permission
None of this requires a technical background. It requires clarity of thought, willingness to be precise, and the humility to notice when what we said is not what we meant.
We are not starting from zero. We are translating something we already know into a context that rewards it more generously than any context before. The person who can write a brief that survives contact with three teams already writes in specification. The person who can delegate without micromanaging already thinks in intent. The discipline is the same. The medium changed. And the medium, for once, is on our side: it rewards the old skill of saying what we mean, clearly and completely, more than any medium we have worked in before.
That is the skill. It is learnable. It compounds. And it has been waiting, patiently, for us to recognize what we already knew.
This is part of “The Hitchhiker’s Guide to the K-Shaped Economy.” Previous: “The New Management” on orchestration. Next: “Trust, But Verify” on evaluation.
