How to find a winning Meta ad in one afternoon, with an AI agent doing the clicking
Most brands do not have a creative problem, they have a loop problem. Here is the afternoon loop we run for our own ads: start from the longest-running ads, pick one angle, launch paused with a target, and read the results the next morning.
Most brands do not have a creative problem. They have a loop problem. The ad gets made in one tool, downloaded, uploaded somewhere else, rebuilt by hand in Ads Manager, and by the time anyone reads the results the person who made it has moved on. So nothing gets learned and next month starts from zero.
Here is the loop we run for our own ads and for the brands using Hermoso. It fits in an afternoon, it costs almost nothing to try, and an AI agent does the boring parts. You can run it with Hermoso connected to Claude, or with any agent tool that can touch your ad account. The loop is the point, not the tool.
Step 1: steal from the ads that have been running the longest
Every brand has competitors, and every competitor has an ad library. The trick almost nobody uses: sort by how long an ad has been running. A brand does not keep paying for an ad for 200 days because they like it. They keep paying because it works.
Ask your agent for the longest-running ads from three competitors in your category. Not the newest, not the prettiest. The oldest. You will usually find the same two or three angles repeated across all of them. That is the market telling you what converts.
Write those angles down in one line each. "Tastes like the real thing, none of the sugar." "The one your dermatologist would pick." "Set it up once, forget it." That list is worth more than any brainstorm.
Step 2: pick one angle and one format, not five
The mistake here is making ten ads. Make three variations of one angle in one format, so that when one wins you know why.
For most product brands the format that wins first is a plain UGC style clip: one person, one product, a real room, talking to camera for fifteen seconds. Not a cinematic film. The film can come later, once you know which angle to film.
Tell the agent the angle, the product and the format. In Hermoso that is one sentence: "Make a serum UGC ad on the dermatologist angle, three variations." It pulls your product photos and brand voice from what it already knows about you, so the ad comes out with your actual packaging in it, not a guess.
Watch the three clips before you do anything else. If a hook makes you look away in the first two seconds, kill it now. It is cheaper to kill an ad in your library than in your ad account.
Step 3: launch paused, with a number attached
This is where most people freeze. Ads Manager has 40 fields and every one of them looks important.
Give the agent a target instead of a form. "Launch these three at $80 CPA, $50 a day, broad targeting, paused." A CPA target does two things: it tells the platform what a win looks like, and it tells you when to stop. Broad targeting is not laziness. On Meta today the creative is the targeting. The ad finds its own audience if the hook is right.
Paused matters. The agent builds the campaign, the ad sets and the ads, and then reads the whole thing back to you before a cent moves. You flip it on. Nothing should ever start spending your money because a model thought that was a good idea.
Step 4: read the results the next morning, then do only one of two things
Do not touch anything for 24 hours. Then ask one question: which of the three is closest to the target?
If one is under target, or clearly trending there, move budget to it and make three more variations of that one. Same angle, same format, different hooks. That is scaling.
If none of them are close, do not tweak the losers. Go back to step 1 and pick the second angle on your list. Angle beats optimization every time. You cannot fix a wrong message by changing the thumbnail.
Step 5: make the loop run without you
Here is the part that makes this compound. Every week the agent pulls the results, writes one line about what won and what lost, and that line goes into the brand's memory. Next week's ads start from that line instead of from nothing.
After a month you have a list of angles that work for your product, with numbers next to each one. After three months you have something most agencies never produce for a client: proof.
Why an agent and not a person
Not because it is smarter. Because it does not get bored. Steps 1, 3 and 5 are the ones humans skip. Nobody wants to scroll a competitor's ad library for an hour, nobody wants to build the same campaign structure for the eleventh time, and nobody writes the weekly note. An agent does all three the same way every time, which is exactly what a loop needs.
What to do today
Pick one product. Ask your agent for the longest-running ads from three competitors. Write the angles down. Make three clips on the top angle. Launch them paused at a CPA you would be happy with. Turn them on. Read them tomorrow.
That is the whole thing. If you want the version where all of that is one conversation, Hermoso is free to start and it plugs into Claude, ChatGPT and the studio at app.hermoso.ai. But do the loop either way. The loop is what works. For the research step in detail, see how to use the Meta Ad Library and why run length is the best free signal.
First published as an X Article on September 9, 2026.