You can't automate what you can't articulate. If you can't describe a task in clear, concrete, objective language, an AI agent (or a new hire) can't do it well, because it doesn't know what you're doing, why, how, or what good looks like. The better you articulate the steps, the better your outputs from AI. That's the articulate economy.
This post expands a 64-second video that goes live on September 24, 2026 at 10am ET. The embed will be added here once it's published.
Key takeaways
- The line is Nicolas Cole's. "You can't automate what you can't articulate." He's someone I consider a mentor, and it's my favorite idea to come out of the AI zeitgeist.
- Proficiency is knowing the steps. If an outcome takes 10 steps to go from 0 to 1, being good at it means you can name all 10.
- Articulation is the handoff. A task you can describe in full can be handed to an AI agent, or to another human, and come back as good as if you did it.
- A good brief has five parts. What you're doing, why, how, what the agent needs to know, and an example of what good looks like.
- Every system we run at VPPA started as a written task. The scheduling, the captions, the clip selection: each one is a document before it's an automation.
The articulate economy is one of my favorite concepts to emerge from everything that's been going on with AI. To explain what it means, I'll quote Nicolas Cole, who always says that you cannot automate that which you can't articulate. I've found that to be true in every system we've built, so here's the idea in full and how to apply it.
What does "you can't automate what you can't articulate" mean?
It means an AI agent can only do a task as well as you can describe it. If the description is vague, the output is vague. If the description is complete, the output can be as good as yours.
Here's how I think about it. If you're very proficient at something, you understand the task very well. Say a certain outcome you want requires ten different steps. The task starts here, it ends there, and to go from 0 to 1 you pass through ten little steps. The better you understand those steps, the better you can articulate them, express them and communicate them using very clear, concrete, objective language.
And the better you can do that, the more likely you are to get very good outputs from AI, because you can explain exactly what you're doing, why you're doing it, how you're doing it, what the agent needs to know, and what an example of good looks like.
Cole put it bluntly in a post on X: "You can't automate what you can't articulate." The rest of that post is a guess about where most AI usage goes: people asking a model, over and over, to do things they don't really know how to do themselves.
What is the articulate economy?
The articulate economy is the idea that, once building and executing get cheap, the people who earn the most leverage are the ones who can describe work precisely enough to hand it off.
For most of history, the person who could do the task and the person who could explain the task were often the same person, and the doing was the valuable part. AI splits them. The doing is increasingly on call. The explaining is the scarce part, and it turns out to be a skill most people never practiced, because they never had to.
I wrote about the adjacent idea in the most valuable skill in the age of AI: everyone now has a builder, so the differentiator is the person who knows what to ask for. The articulate economy is the same observation from the other side. That post is about vision. This one is about process: the ten steps between 0 and 1.
Why do AI outputs come out generic?
Because the instruction was generic. The model fills every gap in your description with the average of everything it has read, and the average is what generic looks like.
When someone tells me "AI is mid," I ask to see the prompt. It's almost always a sentence or two: "write a caption for this video," "summarize this call," "make this better." None of those say what the task actually entails. A person handed that instruction would come back with questions. The model comes back with a guess.
The test I use: could a smart new hire, with no context on my business, do this task from my written instructions alone? If the answer is no, the problem isn't the model. It's that I haven't articulated the task yet.
How do you articulate a task so an AI agent can do it?
Write down the ten steps, then add the five things the agent needs around them. Here's the structure I use, which is the same one Cole describes in the video's source idea:
- What you're doing. The outcome, in one sentence. "An Instagram caption that gets this video found in search and quoted by AI engines."
- Why you're doing it. The purpose changes the decisions. A caption written for search reads differently from one written for a friend.
- How you're doing it. The actual steps, in order. This is where proficiency shows. If you can't list them, you don't understand the task as well as you thought.
- What the agent needs to know. Constraints, rules, things to never do, the vocabulary of your business, the mistakes you've made before.
- An example of what good looks like. One finished, real example does more than a page of adjectives.
A concrete case from our work. Our SEO captions skill is a plain-English document that any agent can follow. It names the outcome, explains why each line of a caption exists, lists the steps, includes the rules (five hashtags maximum, never invent a number, never write in third person), and ends with a full worked example. There's no code in it. Before that document existed, asking an AI for a caption gave us something forgettable. After it, the output is consistent enough that we ship it.
Same with clip selection for our highlights channels. We used to pick clips by feel. The moment we wrote down what makes a clip worth cutting, and what disqualifies one, an agent could do the first pass. Writing that document was harder than picking clips. That difficulty is the point.
Does this apply to delegating to people too?
Yes, and that's the part I find most interesting. The better you can articulate a task, the better you're equipped to hand it off to a robot, to an AI agent, or honestly to just another human being, and have them produce results that are as good as yours.
A person can understand very clearly what the task entails, how to do it, and what good looks like, if you've written it down. A person can also ask questions, which hides bad briefs for a while. AI doesn't ask, so it exposes the gap immediately. In that sense, practicing on agents makes you a better manager of humans. You stop relying on the other party to reconstruct your intent.
Who this is for (and who it isn't for)
This is for anyone who is good at their job and frustrated that AI doesn't seem to be. If you're skilled at something, you are sitting on exactly the asset the articulate economy rewards. You just haven't written it down yet.
It isn't for people hoping the model will supply the understanding. It won't. It can help you draft the description, and it can ask you questions to draw the steps out, but the knowledge of what the ten steps are has to come from someone who has done the work.
Common mistakes when articulating a task for AI
- Describing the output, not the process. "A great caption" is a wish. The steps that produce a great caption are an instruction.
- Leaving out the why. Without the purpose, the agent optimizes for the wrong thing and you can't tell why the result feels off.
- Skipping the example. One real, finished example resolves more ambiguity than any amount of description.
- Assuming the model knows your context. It doesn't know your clients, your rules, or the mistake you made last month. Say it.
- Blaming the model for a vague brief. If a new hire couldn't do it from your instructions, the agent can't either.
- Never rewriting the document. Every bad output is a missing sentence in the brief. Add it and the next run improves.
The better you can communicate, the better you can articulate your ideas, the more leverage you'll have when using AI tools. I'm super bullish on this as the skill to practice right now, because it compounds: every task you articulate once is a task you never have to do by hand again.
Frequently asked questions
Who said "you can't automate what you can't articulate"?
Nicolas Cole, the writer behind Ship 30 for 30 and Write With AI. He's used the line repeatedly in his writing and on X, and it's the core of what he calls the articulate economy.
Is this the same as prompt engineering?
Prompt engineering is one application of it. Articulating a task is the underlying skill: understanding the steps well enough to write them down. A good prompt, a good skill file and a good SOP for a new hire all come from the same ability.
What if I can't list the steps of my own task?
Then you've found the real work. Do the task once, slowly, and write down each decision as you make it. You can also have the AI interview you about the task; the questions it asks are usually the gaps in your description.
How detailed does the description need to be?
Detailed enough that a smart person with no context on your business could do the task from the document alone. If they'd need to ask you something, add that to the document.