Does your AI know how you write copy?

When I started moving my copywriting work from ChatGPT and Claude projects to AI agents, I noticed something I hadn’t expected: the voice wasn’t really aligned with the materials I’d shared.

Before, I’d create a project, put the research and strategy documents there, and prompt the model to help me write. In my experience, it was pretty good at drawing on those materials. I still had to make decisions and edit, but the project gave us a place to work from, and it was easier to iterate on the copy because I was writing section by section alongside AI. The feedback loop was faster.

With agents, I had to think much more deliberately about how to connect everything because once you ask them to do stuff, they get on with it behind the scenes. I’d given the agent a “brain,” basically a knowledge base of project materials and decisions. But those were still disjointed parts unless I put them together.

That was what made it click for me.

Maybe the voice problem wasn’t just about how I’d instructed the agent to write! Maybe it didn’t have a reliable, consistent way to find and use the material I’d already given it.

So I started building a custom skill for this: reusable instructions for how my agents should work through a project’s sources before and during writing.

Why an agent needs more than a folder

If you’ve only used AI in a chat window, an agent takes that a step further. It can use tools to open files, search for information and carry out a sequence of actions, rather than waiting for you to paste everything into the conversation.

You might hear people call the software around the model its harness. That’s the part that runs the model, handles its tool calls and carries information between steps. It helps the agent do work, but you still need to tell it how you want that work done.

Plus, prominent agent tools like Codex have clear coding roots. OpenAI introduced Codex as a software engineering agent, with tasks like writing features, fixing bugs and running tests. Anthropic’s work on long-running agent harnesses took inspiration from how human engineers work: read the project’s progress, choose a feature, implement it and check that it works.

That doesn’t mean these tools can only code. Anthropic explicitly describes its Agent SDK as general-purpose. But a workflow for navigating a codebase doesn’t automatically give you a workflow for using customer research to write a page.

And “act as an expert copywriter” is just vague fluff. It names a role, but it doesn’t explain how a copywriter works day to day, behind their desk, sweating every little detail. And certainly not how you work.

You can tell an agent what your copy should look like, share your positioning and give it examples of how your customers speak. But you also need to teach it how you think through those materials to arrive at the right words for the right people.

What I’m actually doing now before I write copy

When I think about writing a message, I go back to our positioning first. What makes this product different? Which value theme are we writing about?

Then I trace that against the messaging framework: how have we decided to communicate that value, and which messaging pillar does this belong to?

Say I’m working on a section about a product’s features. I’m not just looking for a list of things the product does. I’m going back to the relevant differentiated value theme and the capabilities that make it possible, then deciding which ones belong in this section. To explain the benefits, I’d look at the messaging framework again. And on and on.

It’s a constant back and forth between the positioning canvas, messaging framework and value proposition canvas, where we’ve mapped the customer’s jobs, pains and desired gains. What does this section need to do? What’s the argument? Do we start with a problem or a desired outcome? How do we frame the solution?

This is the reason why I still think a good copywriter should immerse themselves in the market, even when AI can do a lot of the work. Because it’s this kind of work that forms your judgment. The judgment you then need to guide AI.

Copywriting principles and formulas can help me build that argument. But I can only make those choices because I know which piece of material to look at and what I’m looking for in it.

Then there’s the language itself. If a couple of words feel flat or vague, I’ll go back to customer reviews, testimonials or interview transcripts. How do these people actually describe the problem? Is there a more specific, relevant way to say this?

That’s often the last touch, but it isn’t a matter of sprinkling quotes over the copy. The words need to fit the reader and the argument. I want the customer’s way of saying something without taking it out of context or turning a wish into a promise our offer doesn’t support.

All of this happens while I’m writing. I don’t read the documents once, close them and hope I remember enough. Without guidance, an agent can do much the same.

That’s the back-and-forth I wanted the skill to help my agents follow.

What to put in your own copy skill

You don’t need my exact documents or setup. Start by explaining how you use the materials you already have. But I’d include these five things:

1. Define the assignment before choosing the sources.

Who are you writing for, what are you writing, and what should it help the reader do next? Name the project and the approved materials the agent can use. “Search my brain” is too broad if it contains old offers, other clients and ideas you haven’t decided to pursue. And don’t worry, once you guide it, your agent will help you.

2. Give each source a job.

Explain which document governs positioning, which governs messaging, where to check offer details, and where to find customer language and voice examples. Mark what’s approved and what’s still exploratory. Leave out what’s just brainstorming or random ideas.

3. Show the agent how you move between them.

Give it an initial reading order, then explain when to return to each source. Choosing which capability to emphasize? Check the differentiated value. Explaining the benefit? Check the messaging and the customer’s desired outcome. Fixing vague language? Reopen the relevant customer research. Ask it to read the actual passages, not just a summary of what’s in the folder.

4. Tell it what to do when the material doesn’t answer the question.

If the offer document doesn’t confirm what you need to write about, the agent should flag it rather than fill the gap with something plausible. If two sources disagree, it should surface the conflict. You need to make the missing decision before the agent writes as though you’ve already made it.

5. Check its choices, not just its sentences.

Before you approve the copy, have the agent show which sources support the main promise and why it chose that message for this reader. Then check those choices yourself. A headline can sound human and still lead with the wrong value. I have my agent add comments to each piece of copy explaining its choices, so I can follow how it arrived at the copy and double-check those decisions myself.

I was shocked at how much better the copy was and how many fewer rounds of feedback my agent needed when I tested this skill on a rush project I had to get done in one day.

The best part is that I’m still responsible for deciding whether the copy works or not. What I’m trying to make more reliable is the process the agent follows before I even look at the copy.

If you want to try this, take one section of a page you’re writing and talk through what you do before you write it. Which document do you open? What decision does it help you make? What makes you go back to the research?

Put that into your agent’s instructions, then watch whether it actually follows the process.

If anything, it’s a great exercise you can do to reconnect with your own discernment and workflow again.

Discovery

The work between your tools

​Hiten Shah’s piece on why marketing is a good test for AI agents connects directly to what I’m working through above. Your CRM, customer conversations and website can each tell you something different. Someone still has to connect those facts, decide which source matters and carry that context into the next step.

To me, that’s why setting up an agent means explaining how you think and act in your work, not just giving it access to your tools. Hiten adds a useful test: does yesterday’s work make the agent’s next run easier, or are you still rebuilding the context for it?

Faster checks while the agent works

My friend Mike Taylor tested TypeSafe’s Jev for Every, a model that returns structured probabilities rather than chatty answers. One use he explored was checking writing for specific problems, quickly enough to run checks while an agent works rather than only at the end.

It makes me think parts of what we’re doing with the copy context skill could become much faster and more integrated: did the agent find the right material, and does the draft meet the checks we’ve defined? Speed doesn’t settle whether those judgments are right, which Mike is careful about too. But I’m curious about catching problems as the copy develops, instead of saving them all for review.

Resonance

Dan Sullivan, 10x Is Easier Than 2x