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.

1. Get the URL. Higgsfield's MCP server address is https://mcp.higgsfield.ai/mcp (it's on higgsfield.ai/mcp — copy it from the source, never from a spelling you heard out loud).
2. Open Claude's connector settings: click the + under the message box → Connectors → Manage connectors. (On the web app this also lives under Settings → Connectors.)
3. Click Add custom connector, paste https://mcp.higgsfield.ai/mcp, and name it Higgsfield. Click Add.
4. Authenticate. Higgsfield uses account-based login (no API key to paste) — click Connect and sign in with your Higgsfield account. It now appears under your custom connectors.
5. Set the tools to Always allow so the agent isn't halted for a permission click on every single generation, then refresh the tools list. Scroll the connector menu and you should now see Higgsfield's tools available.

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.

• GOAL: Build me a full set of ads — both image ads and video ads — for [OFFER NAME], in a [BRAND STYLE, e.g. 'clean paper / motion-graphic'] style.
• WORKFLOW: 1) Scrape [OFFER PAGE URL] and pull the headline, the proof/results, the imagery, and any photo of me. 2) Read [SCRIPT OR USE-CASE LIST — e.g. a file path or pasted notes] and extract every use case or benefit I'm showing. 3) Using the Higgsfield MCP, generate the image ads and video ads IN PARALLEL — one creative per key use case. Also make one ~60-second motion-graphic video ad in that paper style. Write your own detailed prompts per model; if a generation comes out wrong, rewrite the prompt and regenerate.
• SELF-QA (do not skip): After generating each creative, look at it. Anything with misspelled or garbled on-screen text, a broken layout, the wrong logo, or an off-brand look → move it to a rejected/ folder. Clean, on-brand creatives → move to an approved/ folder. Keep generating until I have at least [N, e.g. 8] approved creatives.
• HANDOFF: Organise everything neatly into a single project folder with image-ads/approved, image-ads/rejected, video-ads/approved, video-ads/rejected, and open it in Finder so I can browse it. Then summarise what you made and what you rejected and why.

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.

• What 'good' looks like in approved/: on-screen text is spelled correctly and readable, the layout isn't broken, your logo/URL is intact, the look matches the brand style you asked for, and the message maps to a real use case from your offer.
• What lands in rejected/: garbled or misspelled captions, a mangled URL, warped faces or hands, the wrong brand colours, or a creative that's just off-tone. These are exactly the failures human reviewers waste hours catching.
• The human still reviews approved/ before anything ships. Model self-QA is best-effort taste — it reliably catches obvious junk, not every brand nuance or a subtle factual slip in the copy.
• The real win: you go from sifting forty files to approving a curated handful. 'Slot machine' becomes 'slot machine with a filter'.

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

• Tools not showing? You forgot to refresh the tools list after adding the connector, or the auth didn't complete. Re-open the connector, confirm it's authenticated, refresh.
• It described the ads but made no files. You ran it in a plain chat window with no filesystem access. Re-run in a session that can write folders (Claude Code or the desktop app pointed at a working directory).
• Credits. Generating a whole campaign in parallel burns Higgsfield credits fast, and 'Always allow' lets it do so unsupervised — start with a small N (e.g. 4 approved) on a real brief before you let it run big.
• Don't trust spellings from a video. Copy the MCP URL (https://mcp.higgsfield.ai/mcp) from Higgsfield's own page, not from anything you transcribed by ear.
• Parallel ≠ instant. Concurrent jobs still queue against the provider's limits; a full campaign is minutes, not seconds. The agent will tell you when it's done — go do something else and come back.

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.

1. Scope — name the one deliverable: '8 approved ad creatives + one 60-sec motion video from my offer page'. Specific and checkable beats 'some ads'.
2. Goal — give the agent that outcome up front in one instruction, including the self-QA bar. A well-specified goal is what lets it run without you babysitting.
3. Workflow — fix the order: scrape → extract use cases → generate in parallel → self-sort. This is your reusable engine.
4. Handoff — define the artifact exactly: the named folders, opened in Finder, plus a written summary of what was rejected and why. The folder is the product, not the chat.

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.

• Part 1 — The always-on operator: give Claude a standing goal and let it work. → guides.kno2gether.com/fable5-payroll-goal
• Part 2 — The bookkeeper: Claude reads your statements and flags the subscriptions to cancel. The computer-use 'cancel' step is review-and-permission gated — it surfaces what to cancel; you click the final button. → guides.kno2gether.com/fable5-payroll-bookkeeper
• Part 3 — This guide: the one-agent ad team that builds and self-QAs your campaign. → guides.kno2gether.com/fable5-payroll-admaker
• Part 4 (next) — Prove-it-or-destroy-it: telling Claude to adversarially tear your idea apart instead of agreeing with you. It's research rigour, NOT financial advice — even when the example is a stock hypothesis, the takeaway is the method (red-team your own thesis), never a trade to make. → guides.kno2gether.com/fable5-payroll-redteam
• Part 5 — Finding clients: Claude researches and drafts outreach. Critical guardrail — it drafts, you send. Every message is human-reviewed before it leaves; the agent never auto-DMs or auto-emails. → guides.kno2gether.com/fable5-payroll-clients