AI UGC prompts

Script or prompt: which AI UGC tools take a prompt

Avatar tools take a script and menus. On Arcads, Creatify’s avatar mode and HeyGen’s avatar videos, you pick an actor and paste the words. There is no scene to describe.

Generative video models take a prompt: Google Veo, Kling, Luma and Seedance, directly or through resellers such as PromptWise.

Between the two sit prompt-to-video modes inside avatar tools. Captions’ Prompt to Video, HeyGen’s Video Agent and Creatify’s Agent turn a written request into a finished talking video. Each bills that request its own way.

A UGC video prompt, field by field

All four guides ask for a person, a place, an action and a camera; Veo and Kling add dialogue and sound, and Seedance sets a rule for dialogue language. This template puts those shared fields in the order most of the guides use.

Subject:  [age and look], holding [the product] at [where in frame]
Setting:  [room], [light source], [time of day]
Action:   [one movement, in the order it happens]
Camera:   handheld phone camera, slight shake, 9:16 vertical, [framing]
Dialogue: [Speaker] ([tone]): "[a line about the product]"
Sound:    [the room's own sound]
Exclude:  studio lighting, perfect composition

Each field, filled once:

• Subject: “a woman in her 30s in a grey hoodie, holding a steel water bottle at chest height”. Luma’s guide says “‘A 28-year-old woman’ gives better results than ‘a person.’”
• Setting: “a small kitchen, morning light from a window on the left”. Luma asks for “setting, lighting, and time of day”.
• Action: “she unscrews the lid, then turns the bottle to the camera”. Seedance’s guide orders each shot as “who + where + doing what + how the camera moves”.
• Camera: “handheld, arm’s-length selfie framing”. Luma’s own example: “Handheld phone camera, slight camera shake, 9:16 vertical format”.
• Dialogue: Kling’s guide tags each line “Character (tone descriptor): ‘dialogue’”; Veo’s example reads “the man in the red hat says: Where is the rabbit?”.
• Sound: “kettle hum, a tap running”. Google separates “Sound effects” from “Ambient noise”.
• Exclude: Google recommends describing what you do not want (“wall, frame”) over instructions such as “no walls”.

The spoken line describes the product. A line that claims the speaker used it reaches the FTC’s fake-testimonial rule, as how to make AI UGC ads sets out in step 7.

Model-specific prompts sit on their own pages: Veo prompts for ads, Seedance UGC ads and Higgsfield prompts for ads.

AI UGC influencer and photo prompts

An AI UGC influencer is one character kept the same across many clips, and every guide read does it with reference images. Seedance’s guide binds each subject to its asset with a tag such as <Subject_N>@<Image_N>, and asks that “the same label should be used” each time the subject returns.

Users report Kling holding one face and outfit across several cuts through its elements and references (4 of 37 sources: Reddit, YouTube).

For stills, image models take the same fields minus sound and motion. Nano Banana makes multi-image composites from reference images, and ChatGPT edits uploaded photos with references. Product stills have their own page: AI product photography prompts.

Prompts that get rejected

Face and person filters reject more than real people, users report. On Seedance, filters reject ordinary photos, AI faces and even cartoon faces (10 of 35 sources: Reddit, Reddit). On Veo, “prominent people” errors hit fictional characters, clients’ photos and users’ own faces (13 of 62 sources: Reddit, Reddit).

Google’s guide states the rule behind it: “prompts that violate responsible AI guidelines are blocked” (prompt guide, Safety filters).

The terms go further than the filters. PromptWise’s terms (11 May 2026) bar outputs that feature a real person or brand without verifiable consent (6.2). Clause 4.7 requires you to say an output is AI-generated where it resembles a real person (rights row). A prompt naming a celebrity or another brand’s product fails on both counts.

What a failed prompt costs

A prompt bills when it runs, whether the clip is usable or not.

Captions charges Prompt to Video 16 credits every 6 seconds, and its help centre says credits are used “regardless of if the projected is exported or saved” (help centre, 21 September 2026).

On Max at $24.99 a month, a 30-second video works out at $4.00 per video, if every credit is used (pricing page, read 2 October 2026). A second attempt doubles it.

HeyGen’s Video Agent costs 40 credits a minute, 120 in Seedance mode (help centre, 17 September 2026). Of 39 HeyGen sources, 4 say the agent ignores instructions and spends far more credits than a standard avatar video (Trustpilot, Trustpilot).

HeyGen Review 2026: Is It Actually Worth It? (5 Things You Need To Know)Simon Crowe

Creatify’s Agent uses a variable number of credits that “cannot be determined upfront” (help centre, 19 August 2026). On the generative models, users report the same pattern:

• Seedance: credits go faster than plan cards suggest (11 of 35 sources: Reddit, Reddit).
• Kling: failed and rerolled takes are charged (10 of 37 sources: Reddit, Reddit).
• Luma: failed clips and agent loops all consume credits, with no estimate first (16 of 36 sources: Reddit, Trustpilot).
• Veo: weak prompt adherence means a usable clip takes several paid retries (13 of 62 sources: Reddit, Reddit).

Luma’s own advice cuts the bill: refine specific elements over regenerating the whole video.

Using ChatGPT or Claude as an AI UGC prompt generator

A chat assistant drafts the prompt; it does not make the video. Paste the template above with your product and audience, and ask for five filled versions.

ChatGPT has a Free plan that writes scripts and hooks (pricing, read 23 September 2026). Its terms assign you “all our right, title, and interest, if any” in Output, and bar you from representing “that Output was human-generated when it was not” (read from an archive capture of 20 September 2026; rights row).

Claude writes copy and briefs and makes no images or video. It has a Free plan, and Pro costs $20 a month (pricing, read 2 October 2026).

Claude’s consumer terms assign you Anthropic’s rights in Outputs, “if any” (section 4, effective 8 October 2025; rights row).