Prompts Are Dead, Long Live Direction
A client showed me her prompt library in March. Two hundred and forty templates, colour-coded by department, exported from a SaaS tool that had charged her team €180 a month for the last eighteen months. The first template began Act as an expert lawyer with twenty years of experience. The second began You are a senior marketing strategist. The third was Pretend you are Steve Jobs.
I asked her what the templates had produced, on a good day, that her team had actually used.
She paused. Some onboarding copy. A FAQ.
The library, in eighteen months, had paid for one FAQ.
This piece is about why that math is not unusual, and what is replacing the discipline we briefly called prompt engineering.
The short window when prompts were the lever
In late 2022, when ChatGPT landed, prompts were the lever. The models were good enough to surprise you, and brittle enough to need careful handling. Add the phrase think step by step and the math improved. Open with a role and the register followed. Drop in three examples and the format locked. There were patterns, and the patterns worked, and a small craft formed around them.
Some of the patterns held real insight: chain-of-thought, few-shot framing, structured output cues. Most of them were superstition that survived because the model was generous to anyone who tried. The prompt whisperer job posting was a 2023 artefact, the way webmaster had been a 1998 one. Both titles were honest at the time. Both faded for the same reason: the underlying system absorbed the trick.
By 2026, the tricks are baked in. Step-by-step reasoning is the default behaviour. Role priming is automatic. Long context lets the model find the structure on its own. The marginal lift you get from a more clever opening sentence is now, in any honest benchmark, negligible. The variance has moved.
Where the variance went
Sit in front of two teams using the same model and one will get an order of magnitude more value than the other. The gap is no longer in the prompt. It is everywhere else.
The variance is now in:
Context. What does the model see before the question? The corpus, the document, the prior decisions, the constraints, the unsaid rules of the house. A blunt prompt with the right context will beat a polished prompt with no context, every time.
Composition. Is one model doing all the work, or are three? Is there a critic in the room? Is there a router deciding which agent picks up the task? A single-agent setup is a soloist; a composed system is an orchestra.
Verification. Who reads the output before it ships, and with what map? The expert’s eye, applied at the right moment, is the actual quality gate, not the model’s confidence, and certainly not the prompt’s elegance. That is the structural argument of The Expert’s Paradox; it does not go away because the prompt is prettier.
Decision. What gets approved, what gets killed, what gets reworked, by whom. This is the human seat. It does not delegate.
None of these is a prompt. All of them are direction.
Direction, defined
Direction is the discipline that replaces prompt engineering. It is the work of shaping the system around the model, not the sentence inside the model.
A director does not type magic words. He chooses what enters the room. He chooses who argues with whom. He chooses what counts as done. He keeps the final say. The prompt itself, in this frame, becomes a small implementation detail: important the way a stage cue is important, but not where the show is made.
Prompt engineering was a craft of insertion: you tried to drop the right keyword combination into the slot and hoped the slot machine paid out. Direction is a craft of composition: you build the room, you populate it, you set the rules, and you stay in charge of the outcome.
The naive chat user still asks what should I type? The director asks what should the system see, who should disagree with the first draft, and how will I know if the answer is wrong?
The questions are not on the same continent.
What a fractional-CIO mandate looks like now
In my practice, almost no client engagement now turns on a prompt. A typical mandate turns on three things.
Building the brief once. Two hours, a dedicated agent, brainstorming the context until I have a structured document that captures the client’s industry, constraints, prior decisions, vocabulary, vendors, sacred cows, and pending questions. That document is reused everywhere downstream. It is the score the rest of the orchestra reads.
Composing the room. For a vendor evaluation: one agent argues for the incumbent, one argues for the challenger, one stress-tests the GDPR and ANSSI angles, one drafts the cost model. They run in parallel, they contradict each other on purpose, they all read the same brief. I sit in the middle and arbitrate.
Putting the senior eye at the exit. Nothing leaves the room until a domain expert (internal if there is one, me if there isn’t) has read the artefact against the brief. The expert is not optional. That paradox is the reason the rollout works at all.
The prompts inside this system are dull. Audit this stack against the brief. List the three weakest assumptions. Argue the opposite of the previous output. None of those would sell as a template.
The value is not in the sentences. The value is in the system the sentences sit inside.
The market hasn’t caught up
The prompt engineering market is still alive in the way some markets stay alive after the underlying problem has moved on. Books are still written, courses are still sold, LinkedIn profiles still advertise the title. The €180-a-month template library is still on the invoice.
When an SMB asks me to help them do AI properly, the first thing I usually do is read their existing prompt library and tell them, with as much kindness as the situation allows, that the library is not the problem. The problem is that there is no context, no composition, no verification, and no decision discipline. The library was a way of buying motion without buying direction.
That misdiagnosis is the part that costs real money. Once the team starts spending its budget on direction instead of on prompt libraries, the same models (sometimes the cheaper ones) start producing work that ships.
The model is rarely the bottleneck. The prompt almost never is. The room around the model is where the work happens, and that room is built by hand, by a human who is willing to direct.
Long live direction
Prompt engineering had a useful life. It taught a generation that words sent to a machine could be designed, not improvised. That lesson is not wasted. It just stopped being the lesson that matters most.
What matters now is whether you can build a room, populate it with agents who disagree, hold the score open while the music plays, and keep your name on the final cut.
A prompt is a sentence you sent to a model. Direction is everything you decided before, during, and after.
The first is over. The second is what the work has always been.
The conductor of an orchestra doesn’t make a sound. He depends, for his power, on his ability to make other people powerful. — Benjamin Zander, The Art of Possibility, 2000