Stop writing prompts. Start directing systems.

Two years ago, the most valuable skill in the room was knowing how to write a prompt. People sold courses on it. There were libraries of the things, traded like recipes. If you could phrase a request precisely enough, the model would do something close to magic, and the phrasing was the skill.

That skill is being automated. The models got good at understanding messy intent, and the tools learned to ask the clarifying questions a good prompt used to answer in advance. The clever prompt — the one you saved in a note and reused — is becoming the thing the system writes for you. This is not a tragedy. It's a promotion.

the work moved up a layer

When execution gets cheap, the value moves to the part that's still hard. With AI, that part is no longer phrasing a single request. It's three things the model still can't do for you: briefing a whole system, directing it as it works, and curating what it produces.

• Brief. Not one prompt — the whole intent. The goal, the constraints, the taste, the things to avoid. A brief is a prompt that knows it's the start of a system, not the end of a request.
• Direct. Steer the work in flight. The first output is a draft of a direction, not a deliverable. Directing is noticing what's almost-right and saying which way to push it.
• Curate. Choose what carries your name. The model will give you ten plausible options. Nine of them are fine. Knowing which one is true — that's the entire job now.

why this is the same story we keep telling

We've written a version of this before, because it keeps being true in new rooms. AI moves into the tools you already use, the cheap mechanical part gets absorbed, and what's left is the part that needs a person: judgement, governance, taste. The prompt-to-system shift is that pattern arriving at the creative layer.

It's also why we've never sold a prompt library as the centre of anything. A library of clever phrasings is a depreciating asset — it ages the moment the model updates. A method for directing systems doesn't age, because it's about you, not the model. When the tool changes, the architecture still works.

what it looks like in your health

This isn't abstract. It's exactly how the AI Health Stack works. You don't write a perfect prompt to a chatbot and hope. You brief a system — Research, Ledger, Protocol — with your biological baseline and your constraints. You direct it: this evidence tier counts, that one doesn't; this fits my week, that's a textbook fantasy. And you curate: the protocol that earns its place is the one you keep. The model proposes a dozen reasonable things. You decide which one is yours.

The person who treats AI as a vending machine for answers gets vending-machine health. The person who treats it as a system to direct gets something that compounds — because the judgement is theirs, and judgement is the part that gets better with practice.

what it looks like in your work

The same shift is rewriting how people build brands, content, and products. A practitioner can now direct one system to produce an entire graphic language, a content engine, a lookbook — in a session, from one brief. The bottleneck isn't execution anymore. It's the brief and the curation: how clearly you can say what you want, and how reliably you can tell the keeper from the filler.

Which is good news for careful people. For years, the careful lost to the fast, because execution favoured volume. When execution is free, volume stops being an advantage and taste starts being one. The thing you spent years building — knowing what's actually good — finally becomes the scarce input.

“Anyone can generate ten options. The value was never in the generating. It was always in knowing which one to keep.”

— On curation

how to start directing instead of prompting

1. Stop hoarding prompts. Keep briefs instead — short documents that capture intent, constraints, and taste. A brief is reusable across models; a prompt is brittle.
2. Treat the first output as a question, not an answer. Direct it. Say what's almost-right and which way to push. The conversation is the work.
3. Build a curation habit. Generate more than you'll use, then cut hard. The discipline of choosing is the skill that compounds.
4. Protect your taste. It's the one input the model can't supply and can't copy. Read widely, notice what's good, and trust the noticing. That's the moat.

Prompting was a phase. A useful one — it taught a generation to talk to machines. But the machines learned to listen, and the skill quietly moved on. The people who'll get the most out of this decade aren't the ones with the best prompts saved in a note. They're the ones who can brief a system, steer it, and know — when ten good options come back — which one is true.