One Agent That Builds Your Ad Campaign — and Throws Out Its Own Junk

Running an ad campaign is strategy, copywriting, art direction, production, and quality control stacked into one role. AI image and video generation has always made the production part feel like a slot machine: prompt, pray, regenerate, repeat. The piece that never worked was taste — knowing which of the forty things you just generated are actually usable. In the walkthrough by Samin Yasar, Claude (running as 'Fable') closes that gap three ways: it has eyes, so it looks at every image and clip it produces and bins the obvious junk; it writes its own deep prompts for whichever generation model it's driving, so you don't need prompt-engineering skills; and it loops on a goal until the output passes its own review. The result is one brief in, a sorted campaign out — not a chat transcript, a folder you open in Finder.

On its own, Claude can plan a campaign and then stop — it has no way to actually render an image or a video. A connector gives it hands. The Model Context Protocol (MCP) is an open standard from Anthropic that lets an agent call an external tool server; Higgsfield publishes one such server that exposes image and video generation across many models (Soul, Nano Banana Pro, GPT Image 2 for images; Seedance, Kling 3.0, Veo, Cinema Studio for video). You wire it in once.

Now you hand the agent an outcome, not a click-path. The pattern is: set a clear goal (the finished campaign), describe the workflow (scrape the offer → generate in parallel → self-sort into folders), and name the handoff (a folder you can open). In the walkthrough this is given with deliberately little direction — the point is to let the agent take the wheel. Two things matter: do it in a session that can write to your filesystem (so it can create real folders), and tell it to generate in parallel so you're not waiting on each ad in series. Below is a complete, reusable prompt — swap the bracketed parts for your own offer.

Paste this into your agent session after the Higgsfield connector is live. Replace the bracketed fields. The self-QA instruction in the middle is the whole trick — keep it.

This is the moment that makes the whole thing feel like an employee. After generating, the agent opens each creative, judges it, and files it — clean ones into approved/, the ones with garbled on-screen text or obvious mistakes into rejected/. In the demo it caught its own misspellings and a creative where the on-screen URL came out as nonsense, and quietly moved them to rejected without being told which specific ones were bad. It works because a model that can view its own output can be told to hold it to a bar and loop until enough pass. You reproduce it with the single SELF-QA instruction above.

The setup is genuinely easy, but a few things trip people up. None are dealbreakers — knowing them up front saves a frustrating session.

Once it works once, write it down so it works every time. Every reliable version of this follows the same four beats — fill them in per offer and you have a recipe you can rerun (or hand to someone) without rethinking it.

This is the third of five build-alongs that put Claude to work on real jobs. Each one stands alone; together they're a small AI workforce. Part 4 is the one that changed how decisions get made.