How to Use Muse AI to Research Winning Meta Ads

No advertiser in their practical mind keeps paying for an ad that isn't converting. So the longer an ad has been running, the more likely it's working.

That one assumption is the whole system. I used Muse, the personal AI agent from Meta, to watch the Meta Ad Library, flag the long-running ads, and break down why they keep running. In this post I'll walk through how it's built, what it finds, where it falls short, and where to grab the master prompts so you can rebuild it yourself.

Why I pointed Muse at Meta's own products

I'm not going to pretend Muse replaced my stack. I use OpenClaw every day, and that's not changing. But I do have a rule for testing any new agent from a big lab.

The best way to use an AI agent is to automate the products and services native to its parent company. The parent company knows its own products best, right? It's the most logical thought.

So what does that look like?

• Muse for the Meta Ad Library, Facebook, and Instagram
• Grok for everything on X
• Codex and the ChatGPT tools for OpenAI's world

That's it. I don't ask a new agent to do everything. I give it one narrow, native job and see how it holds up. For Muse, that job is ad research.

The three scheduled jobs

The system runs on three automations inside Muse [01:33]:

• Daily watch. Every morning it opens every tracked ad and checks if it's still alive. It stays silent unless a winner dies, and then it names it.
• Weekly sweep. On Monday mornings it hunts three to eight new winners and breaks down the hook, the beats, and why each one works.
• Monthly audit. On the first of the month it checks that the playbooks still match reality and creates new ones when a new winning structure shows up.

The output lands in a custom dashboard I built with my Clearmud frontend design skill. It runs natively inside Muse. No separate app, no extra tooling.

What each ad breakdown gives you

For every ad it tracks, Muse works through the same layers:

• The hook. What grabs you in the first seconds.
• The angle. Why this ad keeps running.
• How it's built. The setting, the gear, the look.
• The exact shot list. The nitty-gritty of how to recreate it.
• Master prompts. Ready to hand to any AI, my own motion graphics tool, or a static image generator.

For static ads right now, I think ChatGPT image gen and Nano Banana are the two go-to tools. The master prompts are written so you can drop them into either one.

A live example: The Rundown's "three tips for Claude" ad

Let me show you what I mean. I filtered the library to video ads from the industries I work with, and one of the picks was an ad from The Rundown titled "Here are three tips for Claude" [04:09].

The how-it's-built notes called out clear glasses, headphones, a podcast mic, warm shelves in the background, and a serif title. Great detail. Below that sits the shot list, and below that the master prompts.

The prompts have a slot for brand information. I can fill it with Clearmud branding or a client's branding, and I can save brand kits by name and swap between them. One ad breakdown turns into a template I can reuse across clients.

What the Meta Ad Library gives you, and what it doesn't

The source for all of this is the Meta Ad Library. It's public. It shows every active ad, its start date, and how many copies are live.

What it doesn't show is spend or results. Sadly.

That gap is exactly why I built the system around ad age. Smart advertisers kill ads that don't work. They don't keep paying for a slot that isn't converting. So when an ad has been running for months or years, that's the closest thing to performance data you can get without being inside the account.

One more note: my library pulls Meta ads in the United States only, not worldwide.

Where this works, and where it doesn't

Honest assessment time. For the AI industry, this method is not as effective. Things move too fast, and ads get swapped before longevity means anything.

But when I look at hospitality, construction, and small business, I find great results. Slower-moving industries run the same winning ad for a long time, and that's where the long-running signal is strongest. If you work with local businesses, this is where the system earns its keep.

The playbook: nine formats

Muse identified nine formats as the most common across the winners it found [05:43]. Each one gets its own playbook with master prompts to recreate it across any model. A few of them:

The ugly ad. Modeled on a Frank Kern talking-head ad that has been running since 2020. One person, one phone, one real workspace. It reads like a post from a friend, not an ad, so the scroll doesn't flag it.

The mini doc. A short documentary-style piece.

The free tool. Lead with something useful and let the offer follow.

The guaranteed video sales letter. High trust, high effort, high return. One advertiser runs a 15-minute version. The playbook says to shoot once and deliver two run times: a two-minute paid social cut and a long version for the landing page.

That last one is the pattern I keep coming back to. One shoot, two cuts, two placements.

Get the master prompts for free

I wanted to show this off in two parts. Part one is the walkthrough above. Part two is the prompts, so you can build your very own ad library.

Go to clearmud.ai/resources and log in. It's completely free. It should be near the top, but if you need to search, search "Muse AI" and you'll find every prompt I used. Or screenshot everything and rebuild it your own way. Either works.

I'm not an AI expert. I'm building in public and sharing what works. If you want a video on a specific topic, drop a comment on the video with who you are, what you do, who your target audience is, and what your question is. I'll add a custom video to the queue.

Watch the full build here: https://www.youtube.com/watch?v=zQX3fZtTAAw