30 UGC variants in a day

Thirty UGC variants in a day is a throughput problem, not a talent problem. One product plate, a pack of distinct openings, burned-in captions, and an agent that will actually run the batch. The unit of cost is credits and hours. It is not a day rate, and it is not a dollar figure on this page.

The number 30 is not a brand promise. It is a size that forces the workflow to be a system: if you cannot name the plate, the eight-to-ten hooks, the caption pass, and the kill list, you will spend the day generating twins. 48% of UGC platforms launched 2023–25 are AI-native (archive.com; creator.co). Volume is no longer scarce. Distinct openings are.

This is not a rate card and it is not a fatigue calendar. It is the production day that feeds those. AI UGC is the look you are batching: feed-native, talking to camera, product in hand. A real creator can sit in one or two of the thirty. They should not be asked to be all thirty.

Lock the plate before you generate anything

The body of the ad is the plate: the product footage that does not change. Hands, pour, lid, screen recording, the SKU on a real counter. Film it once, or composite it once from photography you own. Everything else is an opening stacked on top.

If the product is generated, stop. Thirty variants of a warped label is thirty rejection risks. Keep the SKU real. Generate the presenter, the room around the presenter, the reaction cutaways, the captions. The UGC video generator is built for that overlay shape: talking head, product layer, captions from the voiceover. It is the wrong tool for inventing the bottle.

Write the plate brief as constraints:

• SKU, pack size, and the one demo the hands must complete.
• Lighting that matches a phone, not a catalogue.
• No prices, no codes, no other-brand names in the plate (those belong in a hook you might kill).
• Length of the body: long enough to survive several hooks, short enough that a 15–25 second ad still has a CTA.

Do this before coffee. A plate you "fix later" becomes thirty renders you throw away after lunch.

Then write the hook pack on paper, not in the prompt box. Eight to twelve openings that keep one category constraint and change one other variable — the category hook library is the filter. Kill any line that still works on a different aisle. You need about ten survivors to have a chance at thirty honest variants, because each hook will get a small set of presentation changes (who speaks, first frame, caption wording), not a new product story.

Ten hooks × three presentation changes = thirty files. That identity is the day. Twenty hooks × one presentation is how you get twenty near-copies and call it a pack.

The agent pass is routing, not taste

An agent can take a locked plate, a list of hooks, and a caption style, and return a folder. It cannot tell you that hook 4 and hook 9 are the same diagnosis. That is still a person, and it is the first hour, not the last.

Give the agent a job that is executable:

• Presenter: one look, or three looks you actually want to test, named.
• Voice: one register. Do not ask for "energetic and calm" in the same batch.
• Aspect: 9:16 unless a placement is truly 4:5 or 1:1. Do not generate three ratios until a hook wins.
• Caption style: one. Burned-in, safe area respected, first line = the hook.
• Disclosure overlay if the presenter is synthetic: the label is part of the edit, not a later upload toggle. See AI UGC disclosure by platform when you ship, not when you panic.

What you do not give the agent: "make it viral," "try a few vibes," "surprise me." Those prompts spend credits on files you will not run.

Credits are the budget. Different models debit different amounts per generate; the catalog rows list them. A talking overlay plus a caption pass is cheaper than regenerating a full scene for every hook. Prefer overlay on a real plate. Full-scene AI video is for the few hooks that need a new room, not for thirty. If a render fails a logo, a hand, or a mouth, regenerate that shot — do not regenerate the other twenty-nine.

Throughput is mostly waiting and killing. Queue the batch. While it runs, write the kill list for obvious twins from the scripts alone. When files land, watch only the first two seconds of each, mute first, then with sound. Mute-first catches caption twins. Sound-second catches voice twins. Anything you could confuse with another file in the first two seconds is not a variant. Delete it before anyone names it v2.

Thirty honest files from a day is success. Forty files of which twelve are twins is a slower day wearing a bigger number.

Captions are a production step, not a polish step

UGC without burned-in captions is unfinished for silent autoplay. The first caption line is the hook for a large share of views. If you caption after you fall in love with a take, you will refuse to change the line that is not working because the picture is pretty.

Run captions as a batch on the survivors, not on the graveyard. The caption generator and the add captions to video path are the finishing rail. One style across the thirty so the test is the sentence, not the font. If two variants share a first caption line, they are one variant. Change the line or drop the duplicate.

Safe area and timing are the boring failure modes. A hook that sits under a platform UI is not a hook. A caption that lands a second after the spoken line trains people to wait, then leave. Fix timing on the pack, not file by file, unless one take has a different speech rate.

Do not burn prices or codes into the caption pack unless every placement can use them. A listing recut and a paid-social recut should not share a caption file that contains a handle. Split the export: social pack, clean pack. That is still the same day if the picture is locked.

What the person still does after the folder exists

The folder is not the ad account. Someone still has to:

1. Name the test. One variable per file name. hook-invoice-pdf, hook-invoice-pdf-alt-first-frame. Not final.
2. Pick a first flight. You do not ship thirty on day one. Ship four to six that are maximally different. Hold the rest as the cadence queue. Shipping thirty at once is how you learn nothing and fatigue the audience with near-copies.
3. Match rights to placement. A generated presenter still needs a licence story for the product plates. A real creator in two of the thirty needs a usage grant that matches the ads you will run. Volume does not wash this out.
4. Label synthetic media. If a realistic person is generated, the disclosure is on the file. The platform toggle is extra, not instead.
5. Watch the first two seconds in the ad manager preview, not only in the editor. Compression and UI overlay change what "distinct" means.

Human UGC still rates 81% on authenticity against 63% for AI UGC (Kapwing; Billo; creator.co). The thirty-a-day workflow does not close that gap. It fills the volume cell so the human cell can be scarce and good. If all thirty are generated and they all open on "this changed my routine," you spent a day building the 63% pile.

Teams report up to 11× more video per month with AI workflows (OpusClip; Hedra; Pictory). That multiplier is batching and reuse, not a better taste model. The taste is the kill list.

Sign up if the missing piece is the agent plus captions on one plate. Bring the ten hooks. Do not bring "make thirty of something."

FAQ

Is thirty the right number, or a stunt?

Thirty is a size that forces a plate-plus-hook-pack system. If your account cannot spend on six tests this month, produce twelve and keep a queue. The failure is shipping one "hero" UGC ad and calling the rest a backlog you will never cut.

Do I need a different presenter for every variant?

No. Changing presenter and hook at the same time ruins the test. Lock one presenter for a hook family. Use a second presenter as its own test, on the winning hook, later. Casting every file is how a day becomes a casting session.

How do I keep credit spend from becoming a second shoot?

Overlay on a real plate, one ratio, one caption style, regenerate only failed shots. Full-scene generates for every hook is the expensive path. Read the credit cost on the model row before you batch, and stop when the kill list is empty — not when the number on the folder hits thirty.