I built an AI coach that makes me do the work

A big dilemma for every marketer using AI to write copy. Actually, to revise copy:

You still approve it, but if AI chooses the problem, picks the evidence and writes the argument before you get involved, you’ve skipped several opportunities to practice making those decisions yourself.

Reviewing the result takes judgment too. But recognizing a decent answer and arriving at one are different kinds of work.

I don’t want my writing practice to become just an exercise in “good copy recognition.”

I started learning copywriting in 2015, and I still want the deliberate practice: choosing a message, explaining my reasoning, finding the hole in it and trying again. The fact that I can now get a pretty good answer almost immediately is forcing me to revisit my good ole practice.

So, I built an AI coaching skill that makes me do the work.

A skill is just a reusable instruction document. You tell the AI what its job is, which material it can use, when to help and when to shut up. You can save those instructions in a project or attach them when you start a chat.

Mine gives the AI a way to be useful without immediately giving me the answer. It basically adds friction on purpose. Slightly rude if I say so myself, considering who pays the subscription. But that’s the arrangement.

Practice the decision, not just the headline

In one session, the AI gave me a fictional product and customer research, then asked me to choose a message and explain why. I made my choice, tried a headline and let the coach challenge the promise behind it.

My headline suggested that the software could spare an agency an awkward conversation about extra charges. But the product only helped draft the explanation. Someone still had to have the conversation. A small difference in the wording, a big difference in what I was promising.

The coach didn’t just rewrite the line for me. It made me examine the decision behind it. And when its own examples weren’t clear or it judged my answer against information I hadn’t had yet, I pushed back too. Being the coach doesn’t give AI a free pass.

Build your own AI coaching skill

I truly think that reversing the order of operations here helps. Instead of being the AI’s guide, let it coach you for once. Takes some humility, but it also forces you to take a step back, re-evaluate your skills and upgrade them if needed.

You can reproduce my setup with a plain text document together with your agent. Use it for one skill you want to practice, whether that’s choosing a homepage message, questioning a campaign diagnosis or writing a clearer headline.

1. Purpose and boundaries

Start with the job: “Help me practice making and explaining copywriting decisions. Don’t do the exercise for me.”

Keep practice separate from delivery. The coach should start only when you request a session, leave live work untouched and never treat a practice draft as approved copy. If you need to switch to getting something finished, say so explicitly. Sometimes you have a deadline, and turning every task into school would be unbearable.

2. Case selection

Tell the coach which material it can use: a specific work example you have permission to share, a small packet of public research, or a clearly labeled fictional case.

For public research, keep the actual excerpts and links. For fiction, require consistent product capabilities, some ambiguity and missing information. Otherwise the AI can invent a conveniently perfect customer who confirms whatever lesson it wants to teach. Keep any model answer hidden until you’ve tried, and don’t mistake an invented customer’s response for evidence about your market.

3. One bounded drill

Give yourself roughly 20 to 30 minutes and one task. “Based on the research provided, choose which customer problem should lead this page and explain why” gives you something manageable to work on. “Make me a better strategist” gives the AI room to produce a very impressive syllabus you’ll never finish.

Let the coach choose the next step from what you actually answered, rather than assuming you’re a beginner. It should adapt to your level of expertise.

4. Attempt, question, revision, feedback, alternative

Write this sequence into the instructions:

• you attempt the task and explain your choice
• the coach asks one question about the most consequential gap
• you revise
• it gives direct feedback
• only then does it offer an alternative, if useful

That space in the agent’s response is important, because as you might know from working with AI, it’s easy to get “persuaded” by its own “reasoning.” That’s when you’re just reacting to its choice instead of making your own. Ask for stronger hints or a worked example when you’re stuck, and tell the coach to stop the questioning if it isn’t helping. An endless guessing game isn’t deliberate practice.

5. Criteria and independence

Require the coach to explain each criticism using the reader’s needs, the supplied data or the product’s capabilities. It should separate personal taste from an unsupported promise, accept defensible alternatives and assess each decision using only the information available at that moment.

I was surprised when it kept spotting issues in my judgment. It was right!

Record what you noticed unaided, what needed a hint and what remains untested. As you gain experience, try fresh cases with less help. Also require the coach’s own examples to meet the same criteria.

You haven’t practiced independent judgment if the final rule is always “agree with the chatbot.”

6. Session notes

Save the task, your attempt, the hints, your revision and the next useful practice target. Record coaching mistakes separately from yours, so you can both improve.

When you ask to stop, the coach should stop and save the checkpoint in case you want to continue later on with the same exercise and “case study.”

Leave yourself a decision the AI hasn’t made

For your first session, choose one headline and write down why its promise fits the product before asking for feedback. When the coach challenges you, check the evidence yourself. If the line needs to become less exciting to become more honest, rewrite it. If the coach is wrong, make the case.

What we’re doing here is noticing where we’re being misled by our thinking, where we’re avoiding looking at the data, and what we should do differently next time (when it actually matters).

The era of video courses is done, but we can still learn. And we can do it better, faster, and smarter.

Discovery

An argument for doing less with AI

In “Put Some Brain in Your Thinking,” Alberto Romero argues that outsourcing thought to AI can crowd out the thinking we need to do ourselves. His prescription includes doing less, thinking more and stepping away from AI. My spin: configuring AI so it requires an attempt from you before it offers an answer. Worth reading as you build your copywriting coaching skill.

Give your AI the reasons, not just the history

Fred Graver compares Claude to the memory-impaired protagonist of Memento in this essay about working with his AI collaborator. He keeps an explicit context document explaining the bigger picture, the work and how they collaborate, alongside decisions and corrections. Applied to a company knowledge base, I’d want to preserve why we rejected a piece of copy and all important decisions we make, not just which version won.

Resonance

— Dan Sullivan and Dr. Benjamin Hardy, 10x Is Easier Than 2x​