Copywriting with AI: Put a Specialist Agent on Every Step

A page that could carry any company’s name does not persuade anyone. The reader cannot tell who it is for, what gets better if they act, or why this offer is the one to choose. Asking one model to research, draft, proofread, and approve that page makes the problem worse. The sentences sound finished. The argument is still interchangeable.

The work splits cleanly when you treat it like a small desk. One pass gathers. One pass writes in the brand’s voice. Another fixes spelling and grammar without touching the argument. Another checks that the proof is real. Another flags a claim you cannot stand behind. A last pass scores the draft and does not rewrite it. You still decide what gets published.

Give Each Subagent One Job

A large language model (LLM) will do every job you pile into one prompt, including grading its own homework. These passes are subagents. Each has a single job, and each receives the same three facts before it starts.

Name the ideal customer profile (ICP), the unique value proposition (UVP) you can defend, and the one difference a rival cannot paste onto their own page. Artificial intelligence (AI) can assist with the step. It does not own the brief, and it does not get to decide that the difference is close enough.

Hand each subagent a narrow instruction, and refuse any draft that came back from a prompt that mixed these jobs together.

• Research subagent: Find who this is for, what they have already tried, what competitors promise, and which proof you can cite. Do not draft the page. Do not invent a number.
• Writing subagent: Draft in the brand’s voice. The reader is this ICP. The outcome you can defend is this UVP. The difference is one sentence. Do not add proof you were not given.
• Spelling and grammar: Correct mechanics only. Do not sand off the UVP, the examples, or the difference.
• Verification subagent: Check each proof point against its source. Flag anything you cannot find.
• Compliance: Flag a claim you cannot support, a missing credit, and wording you should not publish.
• Judge subagent: Score the draft against the brief. Do not rewrite it. Fail it if the ICP, UVP, or difference is missing.

The balance is the point. The writing subagent is not the fact-checker, and the judge subagent is not a second writer. When one pass starts doing another pass’s job, you are back to a single model approving itself. The ICP, the UVP, and the difference have to be in the brief, or every subagent will invent a generic reader and a generic promise.

Know the Ideal Customer

The first human decision is who the page is for, and who it is not for. That exclusion is where differentiation starts. A page written for everyone can be copied by anyone, because nothing in it depends on a specific customer.

The research subagent gathers the ICP. It looks for the job this person is trying to protect, the problem they already feel, what they have tried, and the words they use for it. It also notes the rival promises, so the difference is a real gap and not a slogan you hope is true. It returns notes. It does not return a headline.

Ask for the audience in two layers, and make the second layer do the work of exclusion.

• Situation: Occupation, the pressure of that role, and the pain that makes them look for a page like this.
• Stage: Whether they are naming the problem or ready to act. The stage decides which part of the UVP belongs above the fold.

If the research cannot name who the page excludes, you do not have differentiation yet. You have a description of a market. Send it back before the writing subagent starts, or the draft will try to serve every stage at once, and the UVP will blur.

Lead with the Value Proposition

People buy the change, not the ingredient list. The UVP is the outcome you can defend for this ICP. Differentiation is the part of that outcome a competitor cannot claim without being wrong about their own product.

The writing subagent drafts the benefit lines from the research notes and the voice brief. It may not promote a feature as an outcome. The judge subagent, later, fails a draft that opens on specifications. Even a business-to-business (B2B) buyer is a person with a number to hit and a mistake they do not want to own. The emotional stake belongs in the UVP when it is true, and compliance should cut it when it is theater.

Keep the outcome and the stake in separate sentences so neither one masquerades as the other.

• Outcome: What is different in this ICP’s work after they use the offer, in their words.
• Stake: What it costs them to keep the current way. That cost matters only if rivals don’t remove it.

Read the lines back with a competitor’s name in place of yours. If they still work, the UVP is a category claim, not your difference. The writing subagent does not get to talk you out of that test.

Write a Headline the Customer Recognizes

The headline is the first place the ICP either recognizes the UVP or decides the page is for someone else. Cleverness any brand could use isn’t differentiation. Specificity is.

The writing subagent proposes several headlines from the same brief. The judge subagent keeps the ones a member of this ICP would recognize as their outcome, and it rejects urgency that isn’t tied to that outcome. You choose when more than one passes. You do not ask the writing subagent to pick the winner, because it will prefer the line it enjoys.

Constrain the headline before you ask for options.

• Recognition: A person in this ICP can tell at a glance that the outcome is theirs.
• Difference: Swapping in a competitor’s name makes the line false, or at least awkward.

A headline that grabs attention while hiding the UVP will attract the wrong reader. The judge should fail that line even when it sounds strong.

Tell a Story Only This Customer Has

A story is memorable when the reader has lived the situation. A generic transformation, with a vague frustration and a tidy result, is interchangeable. Interchangeable is the opposite of differentiation.

The writing subagent may tell that story only from the research notes and the voice brief. The situation has to belong to this ICP. The turn of the story has to be the UVP, not a moral about working harder. It may not invent a customer, a quote, or a result. Verification rejects a story that arrives without a source for any fact. Compliance rejects a story that implies a promise the offer does not keep.

Give the story two boundaries so it cannot drift into a case the research never found.

• Situation: A scene this ICP recognizes, using the pressure the research actually found.
• Voice: The brand’s voice, not a borrowed warmth. The difference should be audible in which details you keep.

If the story still works after you remove the product, you wrote a fable. Put the UVP back in the turn, or cut the story. The writing subagent’s job is the scene. Your job is refusing a scene that any rival could sign.

Keep the Voice and Drop the Jargon

Conversational does not mean vague, and authentic does not mean unedited. The voice brief is a short sample of how the brand sounds, the words this ICP actually uses, and the words you will not publish. Sounding like yourselves is part of differentiation. A polished voice that could belong to the category is not.

The writing subagent holds that brief for the whole draft, from the headline through the last line. Compliance flags jargon this ICP would not use, and any sentence that stretches the UVP into a promise the proof does not cover. Spelling and grammar come later, on purpose. A voice pass that is also a grammar pass will sand off the specific noun that made the line yours.

Put the voice rules where the writing subagent cannot miss them.

• Vocabulary: The ICP’s terms for the problem. Technical words only when that customer expects them.
• Consistency: The same voice on the headline, the story, and the ask. A new tone is a new, weaker difference.

Read the draft aloud. If you cannot hear the brand, the writing subagent followed a generic voice sample, or you never gave it one. The ICP will not supply that voice for you.

Prove What You Claim

An outcome without proof is a wish, and a wish does not differentiate you. The research subagent finds the case, the customer’s words, or the figure. The verification subagent checks that the source exists and that the sentence matches it. The writing subagent does not get to supply a number the research never handed over. If verification cannot find it, the line comes out, even when it flatters the UVP.

I use Grok when I draft so a sampling watermark doesn’t weave through sentences I intend to sign. Which platforms mark text, and why that mark is not authorship, is the subject of Understanding AI Watermarking: Who Owns Your Words? Compliance should flag a draft that would publish someone else’s marked wording as if the byline were the whole story. The UVP still has to be yours. A model can help you say it. It cannot be the proof, and it cannot be the difference.

Separate the kinds of proof so a fluent sentence cannot stand in for a source.

• Customers: Real words from a real customer in this ICP, or no testimonial at all.
• Figures: Use a number only after you’ve verified it matches the source. No rounded cousin of a statistic you half remember.

Proof that belongs to a different customer, or to a different outcome, is decoration. It does not make the UVP more believable, and it does not make the difference sharper. Cut it.

Name the Next Step

A call to action (CTA) is the next step for this ICP, not a button that says the category’s favorite verb. A stranger who is still naming the problem should not be told to buy. A buyer who already understands the UVP should not be sent back to a definition. The difference shows up in the ask: what this person gets for this step, on this offer.

The writing subagent drafts the line from the stage the research found. The judge subagent fails an ask that ignores that stage, and it fails a clever line that hides the outcome. Compliance fails an ask that promises a result the proof section does not support.

Write the ask so the value and the action are the same sentence.

• Action: An imperative this ICP can picture doing next, in their situation.
• Value: The piece of the UVP they receive for taking that step, not a generic invitation to learn more.

If the CTA would fit a competitor’s page, it is not carrying your difference. Send it back to the writing subagent with the stage and the outcome restated, and do not let it invent a new promise to make the button sound larger.

Match the Medium and Ask for the Package

The ICP, the UVP, and the difference stay. The length and the shape change. A short post cannot carry the proof a case study can. It can still say who it is for and what changes. Dropping the difference because the field is small is how every brand ends up with the same line.

The writing subagent adapts the draft to the medium after the brief is stable. I do not stop at the body. I often ask for tags, social updates, a meta description, and a featured image in the same request. The tags should name the customer, the offer, and the difference, not just the format. The description has to say who it is for and what they get, in about two sentences, because that is the line a search result shows. The image should depict that promise. A generic desk doesn’t tell the ICP they are in the right place, and it doesn’t show the UVP.

Search engine optimization (SEO) is why the description is part of the draft, not a chore after publishing. If the search result doesn’t include the ICP and the UVP, the page underneath may not get the chance to make the difference. The judge subagent should score the description and the image brief with the same rubric as the page.

Change the container only after the brief is fixed.

• Short updates: Who it is for, the outcome, and one detail a rival would not choose.
• Long proof: The story, the accepted source verification, and the ask that matches the stage.
• Spoken scripts: Sentences a person can say. The UVP in speech, not a paragraph read aloud.
• Web pages: A scan path from the ICP’s problem to the difference to the CTA.

Changing the medium is not permission to drop the difference, and it is not permission to skip the package around the copy. Tags, the description, and the image are part of the same argument.

Make the Page Easy to Read

A reader in this ICP will not excavate a UVP from a wall of text. Short paragraphs, decision-focused headings, and a bold lead on a list make the difference visible. They do not create it. A clean layout of a generic promise is still generic.

Spelling and grammar run after the argument is stable, as their own pass. The instruction is mechanical: fix spelling, grammar, and agreement. Do not improve the strategy. Do not swap a specific noun for a smoother general one. That noun is often the difference. A sentence that reads more easily and no longer states the UVP is a failed pass.

Format for a scan, then let the grammar pass do only what it was asked.

• Blocks: Short paragraphs and lists so the ICP can find the outcome without reading every line.
• Headings: Each heading says what that section decides about the customer, the offer, or the difference.

A readable page that could still be any company’s page has only solved the layout. Send it back if the grammar pass polished the UVP into something a rival could publish unchanged.

Score the Draft, Then Test It

The judge subagent scores the finished draft against the brief. It does not rewrite. I want a pass, or a named failure: the ICP is missing, the UVP slid back into features, the difference could be swapped for a competitor, a proof point was not verified, the CTA asks for the wrong step, or compliance flagged a claim. A score without a reason is another way for a model to approve itself.

Testing comes after that rehearsal, with real readers. Compare two headlines or two asks that both survived the judge. If the winner changes a claim, it goes back through verification and compliance before you publish it. The judge is not the market. The market also doesn’t give you permission to abandon the ICP the page was for.

Keep the rehearsal and the live test from doing each other’s work.

• Judge: A rubric on the ICP, the UVP, the difference, the proof, the CTA, and the voice. No new sentences.
• Readers: Two versions that already passed, measured on the action you actually want from this ICP.

A line that wins a click and disappoints the ICP has failed the UVP, even if the test looks successful. Send that result back to the brief before you send it back to the writing subagent.

Another pass will not repair a missing brief. If the ICP, the UVP, or the difference is not on the page, stop and write the brief. Then send the jobs back to the subagents one at a time.