The noun gives AI the subject. The verb decides what kind of work it does.
Sometimes I get defensive about the length of articles I write. Not with myself, of course, because I fancy myself as a competent writer, but from time to time I go back and read things I have written a year ago, and I can see the progress in not only how I put words together to shape ideas, but the confidence with which I do that has also grown.
“Well, I try and use all the words,” is the inside joke I tell those that might be listening, but that is only half of the story. I tend to over-write the first few drafts, and then I am rather ruthless in cutting at around draft three. By the time I’ve finished red-lining a draft, the page looks less like a document than a Fincher crime scene: controlled, methodical, red everywhere, and clearly arranged to make a point.
But enough about me.
My Substack - Some Assembly Required - has always been where I work things out in public. It’s where the ideas are longest, sharpest, and least edited for palatability. It also survives the algorithm.
A client sent back a page of copy with one line above it:
“Can you make this better?”
She wasn’t being lazy. She’d already spent the better part of an hour arguing with the thing and had run out of ways to ask for what she actually wanted. I recognized the sentence immediately, because I’ve done this myself — usually late, usually to something I’d stopped being able to see clearly, usually with the quiet hope that the machine would somehow know what I meant even though I didn’t.
I recently published “Return on Inference“. The piece argued that as AI makes intelligence cheap to deploy, organizations will need to become far more deliberate about what they spend it on. Capital allocation shaped the modern corporation. Inference allocation may shape what comes after. Sadly, the algorithm buried it, but I think it’s a good piece and should have been seen by more people.
Since publishing it, I’ve been thinking about the point where a single person, mid-task, decides what kind of intelligence they actually want from the machine, and the choice of words that are used to get there.
Which might mean the next stage of AI fluency isn’t primarily technical. It’s grammatical.
Make this better
“Make this better” is one of the most common instructions we give AI. Better how? More accurate. Shorter. More persuasive. Less embarrassing. Easier to follow. More likely to survive a meeting with Legal. More like the person who’s supposedly writing it. Less like something that was obviously generated in nine seconds.
“Better” isn’t a direction. It’s a small cry for help. Grammatically, “make” is the verb. “Better” has somehow been left carrying the entire brief.
The model will still answer, and that’s part of the problem. AI is extraordinarily accommodating. It will not stop and say you don’t appear to know what you want. It will simply make the writing smoother, the sentences more balanced, the transitions more professional, and the whole thing about fourteen percent less alive. Then it will hand the result back with the confidence of a tailor who has removed the sleeves because you said you wanted something more streamlined.
The model didn’t fail, necessarily. It may have done exactly what you asked.
The nouns and the verbs
Most prompts make two decisions, not one. The nouns define the territory. The verbs decide what happens there.
Customer loyalty is the subject, a noun phrase. You can ask AI to write about it, summarise it, or compare competing definitions of it. You can also ask it to diagnose why loyalty is declining, challenge the assumption that loyalty exists at all, reframe it as switching friction, model how it behaves under price pressure, identify the evidence that would falsify the whole argument, rank the possible explanations, or recommend which one deserves further investigation.
Same territory. Completely different allocation of intelligence.
Ask a model to “write something about customer loyalty,” and it will write something about customer loyalty. This is the prompting equivalent of walking into a restaurant and asking for food. You’ll almost certainly get fed, if a table is available.
The verbs we overuse
Most everyday AI use runs on a small family of production verbs — write, create, generate, summarise, improve, make. These are useful verbs. They are also the verbs most likely to turn an extraordinarily capable model into a very fast content appliance.
Write the email. Create the deck. Generate the headlines. Summarise the meeting. Improve the report.
The return is visible, because the output is visible. There was no email; now there’s an email. The deck took two hours instead of two days. The meeting nobody wanted to attend has become a summary nobody intends to read, which at least closes the loop.
Productivity has occurred. Time saved is real. Tedious work removed is real. Plenty of tasks in professional life don’t deserve a richer philosophical relationship with the person doing them.
But if production verbs are the only verbs an organization gives its AI, it has made an inference-allocation decision whether it recognizes the decision or not. By default, it has decided to spend abundant intelligence making more things.
The verbs we underuse
The more interesting verbs don’t always produce a finished artifact. They create pressure. Compare. Diagnose. Interrogate. Distinguish. Trace. Challenge. Falsify. Stress-test. Reframe. Select. Reject.
Each one asks the model to take a different posture towards the same material. “Summarise this strategy” asks for compression. “Interrogate this strategy” asks where it’s weak. “Identify what would have to be true for this strategy to fail” asks for conditional reasoning rather than confidence. “Argue against the part I’m most attached to” asks the machine to become usefully disloyal — which is a strange thing to want from a tool built to please you, and worth sitting with for a moment rather than rushing past.
These outputs are much harder to place on a productivity dashboard. That may be exactly their value. They’re more likely to stop an organization from producing the wrong thing faster.
Polish is not neutral
I’ve been learning this from my own writing, slowly and with some resistance.
Ask a model to polish a strategy, and it will make the strategy more polished. That sounds tautological until you notice how much the verb has already decided for you. “Polish” assumes smoothness is the objective. It assumes irregularity is a defect — that a blunt transition should become graceful, that repetition should disappear, that a strange sentence should be tidied into a familiar one, that uncertainty should be tightened into authority it hasn’t earned.
Sometimes that’s exactly what a piece of writing needs. Sometimes the rough sentence is the only evidence a person was ever in the room.
So I’ve mostly stopped asking for polish. I’m more likely to ask a model to identify the live edges, the structural pressure points, and the genuine factual risks, and to leave the optional polish alone. Identify, rather than rewrite. Protect, rather than smooth. Pressure-test, rather than improve.
The model is still helping. It is just no longer allowed to quietly assume that sounding professional is the same as being good.
A sequence of verbs is a workflow
The other mistake is believing we need one heroic prompt — the perfect paragraph of instructions that will make the machine leap from an undeveloped thought straight to a finished answer. Occasionally it works. More often, what actually helps is a sequence: map the territory, identify the contradictions, challenge the premise, develop alternatives, rank them, recommend one, and only then draft.
Some of the quality comes precisely from refusing to ask the model to do all of it at once. Research isn’t synthesis. Synthesis isn’t selection. Selection isn’t writing. Writing isn’t judgment. A system capable of all five still benefits from being told, at each step, which one it’s doing right now.
This is roughly how I’ve built my own AI workflows — one pass to search, another to connect, another to apply pressure, and a human, that would be me, to decide what deserves to continue. The workflow isn’t defined by which model does the work. It’s defined by the verbs I choose to use.
Prompting may be the wrong lesson
A lot of organizational AI training is framed around prompting: templates, context, role definitions, output formats, examples, “think step by step,” capital letters when quietly desperate. Some of this genuinely helps. But prompting can just as easily become another software skill — a technique for operating the interface rather than a discipline for deciding what kind of work deserves to happen at all.
This is the failure I am consistently told about by colleagues who work in the HoldCos. Their AI training consists of an FAQ for “logging in” and a catalog of predefined templates.
The more consequential capability is instruction. Can you tell the difference between asking for production and asking for diagnosis? Can you tell when you need more possibilities and when you need fewer? Can you ask for evidence instead of eloquence, and notice the difference when it arrives? These aren’t primarily technical skills. They’re thinking skills that happen to be expressed through grammar.
Verbs are not magic spells
There will inevitably be a list somewhere — thirty-seven powerful verbs guaranteed to transform your AI prompts, download the PDF, save the carousel, become unstoppable by Thursday.
Changing “write” to “interrogate” won’t rescue missing context, unreliable evidence, an incoherent objective, or someone unwilling to judge the result once it arrives. A precise verb can’t compensate for not knowing what problem you’re solving. And the model doesn’t possess some fixed mental mode called “diagnose” that switches on with the correct incantation — verbs remain instructions, interpreted through context, by something that is still, underneath the fluency, predicting what should come next.
The value isn’t a secret vocabulary. It’s the moment of friction the verb forces on the human: what kind of work am I actually asking for? That question improves the prompt before the model has done anything at all.
The organizational verb problem
There’s a larger version of this same pattern. An organization can buy access to a frontier model and still spend almost all of it on verbs that have existed inside productivity software for two decades — write, summarise, format, schedule, extract. It may get a substantial return from that. It shouldn’t mistake extensive AI use for extensive organizational intelligence.
The verbs give away the ambition. Show me the verbs an organization uses with AI, and I can tell you what it thinks intelligence is for. Is the organization using AI to produce, or to question? To accelerate, or to reconsider? To generate more options, or to work out which option should die? To automate the existing process, or to ask why the process still exists?
This is inference allocation happening at the level of a single sentence. Every prompt spends intelligence against an instruction. Multiply that across thousands of employees and millions of prompts, and those small grammatical choices quietly become an operating model — one nobody voted on, assembled entirely out of defaults.
A company dominated by production verbs will get faster production. That’s genuinely worth having. A company that also learns to reach for analytical, adversarial, imaginative and judgment verbs may start getting something else: not simply more answers, but better questions about which answers are worth having in the first place.
Check your verbs
Next time AI hands you something polished, plausible and quietly unhelpful, don’t immediately rewrite the prompt, swap the model, or decide the machine has no taste. Look for the verbs first.
Did you ask it to write when you needed it to question? To summarise when you needed it to distinguish? To improve when what the piece actually needed was a diagnosis?
The model may have done exactly what you asked. That may be the whole problem — and it’s a more useful one to sit with than “the AI isn’t very good.”
