Hey, it's Cindy ๐ฑ You commented AGENCY, so here is the whole setup: two connectors, six prompts, and the weekly rhythm that turns one product link into a batch of ad creative.
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)
๐ก Read this before you start, it will save you an argument with the internet. This is two separate jobs wearing one coat. Claude plus the image and video models makes the creative. Claude plus Meta's own ads connection is what touches your ad account. They are different connectors with different logins, and no single connector does both. Anyone telling you one chat generates a video and silently launches it live is skipping the part where you approve the spend, which is the part you should never skip anyway.
the shape of it
What you are actually building
An ad creative engine is not one clever prompt. It is a small assembly line with four stations, and the whole point is that you can run it again next week without rethinking it.
1. The brief Claude reads your product page and writes the angles, audiences and hooks before a single image exists. Skipping this is why most AI ad batches look pretty and convert badly.
2. The generation Product shots, UGC-style talking clips, and short video variants, generated in batches from the brief rather than one at a time.
3. The selection Claude ranks what came back against the brief, kills the weak ones, and writes primary text, headlines and descriptions for the survivors.
4. The test The winners go into your ad account in a clean, named structure so that next week you can tell which angle actually worked, not just which ad spent money.
before you start
What you need, and what it costs
๐ฉโ๐ป A Claude account. Custom connectors work on Free, Pro, Max, Team and Enterprise, with one important catch: Free accounts are limited to a single custom connector. Since this build uses two (creative, then ads), a Free account will run the creative half only. That is stated in Anthropic's own custom connectors guide. Plans and prices: claude.com/pricing.
๐ฌ A paid Higgsfield account. This is the connector that gives Claude direct access to the image and video models, and it needs an active paid subscription, so treat this as the real cost of the build rather than a free trial you can stretch. It runs on credits: in their own words, "Higgsfield tools use the same credit system as the Higgsfield platform. Each generation costs credits based on the model and resolution" (higgsfield.ai/mcp). Two things to know before you batch: video costs many times what a still costs, and one-off credit top-ups expire after 90 days, so buying a big pack "for later" is how people quietly waste money.
๐ A Meta Business account you control. Only needed for the fourth station. You want admin access to the ad account, because the connection is authorised at the business level, not from your personal profile.
๐ One product or landing page URL. The whole engine feeds off this. A live page with real copy, real photos and real reviews gives Claude far more to work with than a paragraph you type from memory.
๐ฑ If you only do one half, do the creative half. Generating and ranking a batch of variants is where the time actually goes. Uploading to Meta takes minutes by hand, and doing it by hand for the first month is a genuinely good idea while you learn what your account structure should look like.
connector one
Connect Claude to the image and video models
This is the connection that turns Claude from a writer into a studio. Once it is on, Claude can call the generation models directly and hand you finished files instead of describing what it would make.
๐ Make your account first. Sign up at higgsfield.ai and stay logged in in that browser tab. The connector authorises against this login, so having it live saves you a detour later.
โ Open Claude's connector settings. Go to Customize, then Connectors, then click the + and choose Add custom connector. Direct link: claude.ai/customize/connectors.
๐ Paste the server URL. The address you paste is https://mcp.higgsfield.ai/mcp. Leave the advanced OAuth fields empty, they are for connectors you build yourself. The setup page is at higgsfield.ai/mcp if you want the vendor's own walkthrough open beside you.
โ Click Add, then connect and approve. Claude will send you through a sign-in and permission screen. Approve it, and the connector shows as connected in the same list.
๐งช Test it in one line. Start a new chat and type: "List the image and video models you can reach through my connected generation tools, and tell me the credit cost of each." If it answers with real model names and numbers, you are live. If it says it has no such tool, reopen the chat, since connectors attach when a conversation starts.
๐ง Pasting the URL into a browser looks broken, and that is fine. If you open https://mcp.higgsfield.ai/mcp in a tab you will get an authorisation error. That address is a machine-to-machine endpoint, not a website. Claude is the one that is meant to talk to it.
station one
The brief prompt, run this first
Generating before briefing is the single most common way these batches go wrong. You get forty beautiful images of the same idea. Run this first, in a fresh chat, and everything downstream inherits the thinking.
๐ง build the creative brief
Read this product page and write me a creative brief for a paid social test: [paste your URL].
Start by telling me back, in plain language, what the product actually is, who is clearly being sold to, what the strongest proof on the page is (reviews, numbers, guarantees, named credentials), and what objection a first-time buyer would have. Then give me five distinct angles to test. An angle is a reason to care, not a slogan, so make them genuinely different from each other: for example a problem-first angle, a proof-first angle, a comparison angle, an identity angle, and an objection-handling angle.
For each of the five angles give me: the one-sentence promise, the specific audience it is aimed at, three hook lines under twelve words, and the single visual idea that would carry it. Flag any claim you are inferring rather than reading directly off the page, because I am not putting an unverifiable claim into a live ad. Keep the language plain and specific, no marketing filler, and do not use em dashes.
๐ก Read the brief properly and argue with it. Delete the two weakest angles yourself. Three strong angles generated deeply beats five shallow ones, and the deletion takes ten seconds now versus a week of muddy test data later.
station two
The three generation prompts
Run these in the same chat as the brief, one after the other, so Claude keeps the angles in context. Each one produces a batch, not a single asset.
Product stills first, because they are the cheapest to generate and the fastest to judge:
๐ธ product shots batch
Using my connected generation tools, create product images for angles 1, 2 and 3 from the brief above. I want three images per angle, so nine in total.
For each image, write the prompt yourself from the angle's visual idea, and make the three within an angle genuinely different setups rather than three crops of the same shot: vary the surface, the lighting direction, the props and whether a hand or person is in frame. Every image must be shot in portrait 4:5 or 9:16, keep the product itself accurate to the reference images on the page, and leave clear negative space in the upper third where a headline will sit later.
Before you generate, show me the nine prompts as a numbered list and tell me the total credit cost so I can approve it. After they render, show me each image next to its angle number and prompt so I can tell which is which.
Then the UGC-style clips, which are the ones that usually carry cold traffic:
๐ฅ UGC-style ad batch
Now build UGC-style video ads for the two strongest angles in the brief. Two variants per angle, so four clips, each between six and ten seconds.
Write me the script for each one before you generate anything: a spoken hook in the first two seconds, one specific benefit with a concrete detail rather than an adjective, and a plain closing line. The person should sound like a real customer talking to a friend, not a voiceover artist, so keep sentences short and let them be slightly imperfect. Describe the setting as somewhere ordinary and well lit, in portrait 9:16.
Show me the four scripts and the total credit cost for approval first. Once I approve, generate them, and flag any clip where the lip sync or the hands come out wrong so I do not have to hunt for it myself.
And the wider video variants, which is where the volume comes from:
๐๏ธ video variant batch
Take the single strongest angle from the brief and build me six short video variants from it, each five to eight seconds, portrait 9:16.
Hold the core promise identical across all six and change exactly one variable per variant so the test is readable: variant 1 changes the hook line, variant 2 changes the opening shot, variant 3 changes the pacing, variant 4 changes the setting, variant 5 leads with the proof instead of the problem, variant 6 opens on the product in use rather than on a person. Tell me in one line which variable each one is testing.
List the six prompts and the credit cost before generating. When they are done, name the files so the angle and the tested variable are both in the filename, because I will be matching them to performance data later and I do not want to be guessing in three weeks.
โ ๏ธ Approve the credit cost every single time. Batching is what makes this powerful and it is also what makes it expensive. Asking for the cost before generation is not caution theatre, it is the difference between a fifteen dollar test and a very memorable afternoon.
station three
Rank the batch and write the copy
You now have roughly twenty assets. This is where most people stall, because judging your own batch is genuinely hard. Hand the judging back to Claude with a rubric, then overrule it where your taste disagrees.
๐ rank and write
Score every asset we just generated against the brief and give me a ranked shortlist.
Score each one from one to five on four things and show the scores: does the first frame stop a scroll, is the promise legible within two seconds, is the product represented accurately, and does it match the angle it was built for. Anything scoring three or below on accuracy gets cut outright, whatever else it scored, because a flattering-but-wrong product shot is a refund waiting to happen.
For the top eight survivors, write the ad copy: primary text at around ninety words, a headline under forty characters, and a description under thirty characters. Match each piece of copy to the angle its visual was built for rather than writing one generic set. Then give me a plain table with columns for filename, angle, tested variable, headline and primary text, so I can hand it straight to whoever is uploading.
connector two
Connect Claude to your ad account
This is a completely separate connection, and since July 2026 Meta hosts an official one. Their ads MCP server gives a compatible AI agent permission-scoped access to your ad account, covering comprehensive reporting, ad creation and management, catalogs, signals and datasets, A/B tests and conversion lift studies, and activity logs. It was opened to any app in July 2026.
๐ข Check your access first. The ad account has to sit inside a Business account you administer. If someone else owns it, you need them to grant you access before any of this works, and that request can take days, so send it now.
โ Add it as your second custom connector. Same path as before: Customize, then Connectors, then +, then Add custom connector. The server address you paste is https://mcp.facebook.com/ads. This is also the exact point where a Free Claude plan runs out, since Free allows one custom connector total.
๐ Authorise one ad account, not all of them. When it asks which assets to grant, pick the single ad account you want Claude working in. Broad permissions are the thing you regret, and narrowing them later is more annoying than choosing well now.
๐ Stay read-only for a week. Ask for reports before you ask for actions: "show me last 7 days by campaign", "which ad sets are below a 1% link click rate", "break spend down by creative". You learn how it talks about your account before it can change anything in it.
๐ ๏ธ Prefer the token route? It still exists. If you would rather use a system user and an access token, or drive it from the command line, I wrote that walkthrough separately: connect your Meta Ads account to Claude.
โ ๏ธ The seam between the two halves is manual, and anyone telling you otherwise is selling something. The generation connector has no Meta tools on it at all. Not one. It ends at a rendered file. Meta's ads tools can build campaigns, ad sets and ads, but getting a freshly generated video into the account still means uploading it into Meta's media library yourself. Budget ten minutes per batch for that and stop waiting for the loop to close on its own.
โ ๏ธ Set budgets by hand for the first month. Let Claude build the structure, name things, and write the copy. Type the daily budget in yourself. An automation error in a document costs you a rewrite, and the same error in a campaign costs you actual money overnight while you sleep.
Once the videos are uploaded and the connection is live, this is the staging prompt. Notice it stops before launching.
๐ stage the test
Using my connected Meta Ads tools, stage a creative test in ad account [your ad account ID] using the shortlist and copy from the table above. I have already uploaded the video and image files into the account's media library, so reference the existing assets rather than trying to upload anything.
Build one campaign, one ad set, and one ad per shortlisted creative. Name everything so the angle and the tested variable are readable from the name alone, using the pattern ANGLE_VARIABLE_DATE, because in three weeks the names are the only thing standing between me and a spreadsheet I cannot interpret. Put every ad in the same ad set so they compete on equal footing rather than fragmenting the budget.
Create everything in PAUSED state and do not set a budget, do not set a schedule, and do not publish. When it is staged, give me a checklist of exactly what you created with the IDs, plus anything you could not complete, so I can review it in Ads Manager, set the budget myself, and turn it on.
station four
The weekly rhythm that makes it an engine
One good batch is a nice afternoon. The engine part is the repeat, and Meta's own published guidance points the same way. In their piece on creative diversification they recommend it as "a foundational best practice", and they draw a sharp line worth internalising: "Creative iteration might produce two ads with identical visuals, but different text CTAs, while creative diversification would generate two distinctly different pieces of creative." Volume of near-identical variants is not the goal. Genuinely different creative is.
๐๏ธ Monday, read last week. Ask Claude which angle won, not which ad won. The ad is an instance, the angle is the reusable insight, and only one of those two things tells you what to make next.
๐จ Tuesday, generate. Re-run the generation prompts, but seed them with the winning angle and one fresh angle. Keeping one experiment in every batch is what stops you optimising yourself into a corner.
โ๏ธ Wednesday, rank and write. Same ranking prompt, same rubric. Consistency in the rubric is what makes week five comparable to week one.
๐ Thursday, stage and launch. Stage paused, review in Ads Manager, set your own budget, turn it on. Kill last week's clear losers in the same sitting so the account stays legible.
๐งพ Friday, write it down. One paragraph in a running document: what you tested, what won, what you believe now. This file becomes the thing you paste into Claude next month, and it compounds faster than any prompt.
๐ฑ An honest note on the numbers people quote. "Thirty variants a week" is a description of your process, not a promise about your results. What volume reliably buys you is a faster read on which angle is right. It does not buy you a good offer, a working product page, or an audience that wants the thing. If those are broken, more variants just help you find that out sooner.
the traps
Four things that will bite you
Identical variants If your six "variants" change three things each, no result is readable. One variable per variant, always, or you are just making pretty noise.
Unnamed files Generate forty assets with default filenames and you will never match them to performance. Naming is the boring habit that makes the whole loop worth running.
Invented claims Generated copy will happily promise things your product does not do. Check every number, guarantee and superlative against your own page before it goes live.
Credit burn Video is where budgets vanish. Approve costs before each batch, and generate stills to test angles before you spend video credits proving one.
all the links
Everything you need, in one place
Generation connector: higgsfield.ai, and the server URL you paste into Claude is https://mcp.higgsfield.ai/mcp
Want to build the whole system, not only the creative half? Inside AI System Club I teach people to build AI systems that actually run: content engines, ad engines, and the automations underneath them. Doors open soon and founding pricing dies at close. ๐ฑ Join the waitlist (link in my bio too).
the waitlist is now open
AI System Club is launching soon
Build AI systems with me through recorded lessons you can follow at your own pace. Join the waitlist for first access and the founding price.
btw this list gets first dibs the moment doors open :)