The work happened before the prompt — Editorial Intelligence

AI has made parts of my work ridiculously fast. The speed is real. Where it comes from is widely misunderstood.

I’ve been thinking about why AI has suddenly made some of my work feel ridiculously fast.

Today I can start with a half-formed observation, interrogate it against things I’ve already built, connect it to previous research and frameworks, test the argument, develop it into a long-form piece, turn that into a LinkedIn post and update the underlying methodology.

That can happen in hours.

A few years ago, parts of that process could easily have taken days.

It’s tempting to look at that and conclude: AI did the thinking.

I think that gets the causality backwards.

The speed happens at the end of the process.

A lot happened before the prompt.

Years of writing. Customer interviews. Research projects. Conversations with subject-matter experts. Editorial decisions that worked. Editorial decisions that didn’t. Seeing how organisations actually behave. Trying to translate complicated things into something people understand. Learning which evidence is useful and which is noise. Working on narratives over months rather than treating every article as an isolated object. Experimenting with AI search. Building frameworks. Changing my mind. Recording observations when something didn’t quite fit.

And most importantly: saving enough of that thinking that I can find it again.

So the workflow isn’t:

prompt → AI → idea

It’s closer to:

experience → evidence → observation → interpretation → accumulated thinking → stored context → AI retrieval and connection → new thinking

That distinction matters, because I think we’re currently overestimating the importance of the prompt and underestimating everything that happens before it.

The prompt is not the input

There is an enormous amount of discussion about prompt engineering.

What words should you use? What instructions produce the best result? Which model gives the strongest output?

Useful questions.

But somebody else could copy one of my prompts exactly and they wouldn’t necessarily get the same result.

Because the prompt isn’t the whole input.

The real input is everything surrounding it.

The research. The frameworks. The examples. The mistakes. The previous arguments. The contradictory evidence. The language customers actually use. The things you noticed six months ago and decided were worth keeping. The things you’ve learned by doing the work.

In other words:

You can’t prompt your way around having nothing to draw from.

This is one reason I’m increasingly interested in organisational memory rather than content production.

Most companies already generate enormous amounts of potentially useful knowledge — customer conversations, sales calls, research, product discussions, events, support interactions, executive thinking, performance data, internal debates.

Much of it disappears. Someone remembers part of a conversation. Another person saves a presentation somewhere. Research gets turned into one campaign and then effectively dies. An expert explains something brilliant in a meeting and nobody captures it.

Six months later the content team sits down and asks: what should we write about?

So they start again.

I’ve argued separately that this is what an evidence system exists to prevent, and that content work only becomes valuable when it compounds. What has changed is the economics.

AI raises the reuse value of work already done

Generative AI makes it far cheaper to retrieve and reconnect previous thinking.

It becomes practical to ask: where have we seen this before? What evidence do we already have? Does this new observation support one of our existing arguments, or contradict one? Which customer conversations are relevant? What did we believe about this six months ago? Which ideas have appeared repeatedly but never been developed?

AI is genuinely useful here.

But notice what it is doing.

It isn’t replacing the original work. It is increasing the reuse value of that work.

A customer interview no longer has to produce one customer story. A research project no longer needs to end when the campaign finishes. A conversation doesn’t have to disappear when the meeting ends. A good observation can stay available to collide with something you learn six months later.

That changes the value of doing deep work in the first place, because the return on thinking compounds. An interview generates evidence. The evidence contributes to a narrative. The narrative influences another project. New evidence challenges it. The position gets stronger. Later, an AI system can retrieve the whole chain when a new question appears.

You aren’t beginning from zero.

You’re beginning from everything you’ve already learned.

I’ve written elsewhere about what it feels like when that system gets dense enough to surprise you. The point here is narrower and, I think, more practical: whatever that system appears to do, it is running on work somebody already did.

Two hours, or ten years?

This creates a problem for how we measure the work.

Industrial-era productivity measures the visible labour. How many hours? How many assets? How many words? How many tasks completed?

AI makes those measures increasingly unreliable.

Someone might spend two hours producing something that previously required two days. That doesn’t mean two hours of thinking created it.

There may be ten years sitting behind those two hours.

This is a version of a problem I’ve looked at before in a different setting. The hidden hours research found accountants doing substantial work that fell outside the formal boundaries of their role — real, valuable, and invisible to the systems that measured their output.

AI creates the same gap in knowledge work, but pointing the other way. There, invisible work made the visible hours misleadingly low. Here, invisible accumulation makes the visible hours misleadingly fast.

In both cases the measurement captures the wrong thing.

The difficult work has moved. Away from producing everything manually, and towards noticing, learning, capturing, organising, questioning, connecting, judging — and deciding what deserves to be remembered.

None of that appears in a time sheet.

What have you built for the prompt to act on?

This is also why I don’t think the advantage will simply belong to whoever becomes fastest at generating things.

Generation becomes cheap.

The greater advantage may belong to people and organisations that become very good at accumulating useful context.

Do strong work. Capture what you learn. Preserve the evidence. Develop positions rather than isolated outputs. Record uncertainty. Keep the contradictions.

Make previous thinking retrievable.

Then use AI to interrogate all of it.

The prompt still matters.

But increasingly I think the better question is not how to phrase it.

It’s this:

What have you built for the prompt to act on?

I’ve followed that question further in Context is capital, which argues that accumulated context behaves less like an archive and more like productive capacity — with the maintenance problems that implies.

Because perhaps one of the biggest productivity breakthroughs AI gives us isn’t that it can think instead of us.

It’s that work we already did no longer has to disappear.