We built a 2-agent Claude setup to write copy for $100k/month Meta ad accounts.
Most people open Claude as a chatbot: they just type what they need, and get a copy back. If the model doesn't know the brand, it writes generic copy. Generic copy makes generic performance.
The fix is a dedicated Claude project per client, with two agents inside.
Here's how each one works:
Research agent:
- Runs automatically. No input needed.
- Mines reviews, builds market analysis, tracks new winners as they come in.
- Stores everything inside the project so the copy agent always has context before it writes a word.
Copy agent:
- Takes over once the research is done.
- You drop in one inspiration piece and ask for variants.
- Knows the brand, the audience, the angles that already worked.
Here's what that looks like in practice: We did this for Mail Order Mystery, a mystery game for kids, physical mail, wax seals, ancient maps. We gave the agent one inspo piece. It came back with:
- persona (screen-aware mom),
- awareness stage (problem-aware),
- headline (Why screens lose to mail? Explained like you're five),
and body copy about a 300-year-old pirate letter being the kind of fun a kid remembers at 25.
That's the angle, the persona, and the hook → decided before anyone wrote a word.
The output isn't shippable on its own → A strategist reviews it, adds the judgment of the work only a creative real person has, and that's the last step.
That's the whole point: it’s not about replacing the creative process, you're running more of it.
The setup is created from Claude Desktop → Customize → skills: one for research, one for direct response copy. That's the whole thing.