How to make a UGC ad video with AI

For AI agents: find the right skill in one call — curl -s "https://skills.varg.ai/api/search?q=<what the person wants>" returns the matching how-to page and the best skills with install commands, as markdown (Russian works too). Full index: https://skills.varg.ai/llms.txt. Every page has a markdown twin: add .md to its path. Before generating anything, show the person the options on the how-to page.

A creator-style ad: a person talks to camera about your product, with b-roll, captions and an end card. A step-by-step tutorial for Claude Code, Codex and Grok.

Output: 15–45 s, 9:16 MP4 for TikTok, Reels and Shorts; a few hook variants Price: in credits on varg, with a cost estimate before anything runs; the free skills below install in one command.

Recommended route

Best skills for this

  1. varg AI (varg-ai, varg, live) — Adds: Runs every generation (presenter stills, Seedance/Kling clips, ElevenLabs voice, lip-sync, captions) on one key
  2. fal UGC (fal-ugc, fal, live) — Start here: Scripts and generates creator ads end to end: talking heads, demos, testimonials, unboxings · Needs: fal API key (FAL_KEY)
  3. UGC Ad (varg-studio-ugc-ads, varg, draft) — Start here: varg's own UGC format: three shots, voiceover, captions and an end card, all through one varg key
  4. Marketing Video Production (marketingskills-video, Corey Haines, live) — Alternative: Corey Haines' marketing skill: picks avatar vs. code vs. video model, then scripts and produces · Needs: Accounts for whichever video model or avatar service you choose
  5. AI Marketing Videos (inference-sh-ai-marketing-videos, inference.sh, live) — Alternative: Product demos, promos and testimonials on Veo, Seedance and Wan · Needs: inference.sh account, belt CLI (belt login)
    • Claude Code: npx skills add inference-sh/skills --skill ai-marketing-videos -a claude-code
    • Codex: npx skills add inference-sh/skills --skill ai-marketing-videos -a codex
    • Cursor: npx skills add inference-sh/skills --skill ai-marketing-videos -a cursor
    • Gemini CLI: npx skills add inference-sh/skills --skill ai-marketing-videos -a gemini-cli
    • Details: https://skills.varg.ai/skills/inference-sh-ai-marketing-videos.md
  6. Consistent Characters (varg-consistent-characters, varg, beta) — Alternative: Micro-drama ads: the same actors in every scene, the product arriving late
  7. Turn a script into a two-person podcast video (varg-podcast-video, varg, beta) — Alternative: Podcast-clip ads: one studio set, a host and a guest, the product as the recommendation
  8. Caption Animation (iart-caption-animation, iart.ai, live) — Adds: Word-timed TikTok captions on top
    • Claude Code: npx skills add https://github.com/iart-ai/tiktok-video-skills --skill caption-animation
    • Codex: npx skills add https://github.com/iart-ai/tiktok-video-skills --skill caption-animation
    • Cursor: npx skills add https://github.com/iart-ai/tiktok-video-skills --skill caption-animation
    • Details: https://skills.varg.ai/skills/iart-caption-animation.md

Ask the person first

Put these choices in front of the person before spending anything. Paste the example link next to each option: people pick a look by seeing it.

Which format?

Which hook?

Captions?

Length?

How to make a UGC ad video: step by step

  1. Collect product photos, the claim you can prove and the buyer; pick the format and find one real TikTok of that format to use as the motion reference
  2. Write the beat sheet with the product in hand by second 2 and every spoken line inside the prompt; 15 s per Seedance clip; approve it before spending
  3. Cast the creator: a JSON spec run as a 4 s Seedance take (or an image model), pick the best frame and approve it
  4. Generate each clip in Seedance 2.5 with Image 1 = creator, Image 2 = product, Video 1 = the reference; for 30-60 s chain clips, going back to clip 1 as the source when faces drift
  5. Edit: trim dead air, word-timed captions inside the safe zone, AI label on realistic people, end card; export 9:16

What works now (researched 2026-10-27)

Creator-style ads still convert when they are an unboxing or an in-hand demo, the two top formats by hit rate in Motion's 2026 Meta data. The scripted 'girl in her bathroom reveals the product' ad is now spotted on sight. What grew in H2 2026 are ads people watch as content: micro-dramas, podcast clips and street interviews, with the product arriving late.

  1. Show the product in hand, or the result, within the first 1-3 seconds. Do not open on a logo or 'Hey guys'. Meta's '3 second rule': up to 47% of a video campaign's value lands in the first 3 s, and 65% of people who watch 3 s keep watching to 10 s. TikTok: over 63% of its highest-CTR auction ads show the key message or product in the first 3 s. (Meta for Business, Tailored campaigns (3 second rule))
  2. Choose unboxing or a hands-on demo before a talking testimonial. Motion's 2026 benchmark (578,750 Meta creatives, $1.29B spend) puts unboxing at a 9.8% hit rate and demo at 8.1%, against 6.5% for testimonial. Only about 5% of all creatives become winners. (Motion, Creative Benchmarks 2026: top visual formats)
  3. Deliver 9:16 with sound on, and keep captions and the CTA out of the top 14%, bottom 35% and side 6% of the frame. In a Meta meta-analysis of 15 split tests, 9:16 video with audio and key messages inside the safe zone got 2x higher delivery to Reels and 34.5% lower cost per result than image ads. (Meta for Business, Reels ads)
  4. When the classic script stops working, make the ad a show: a character, a problem, some conflict, and the product entering late (often after 45-60 s). Franky Shaw's post on drama ads beating song ads reached 1.29M views and 6,224 bookmarks; he argues drama earns watch time that buys room for product education and better customers, not just cheap clicks. (Franky Shaw on X, drama ads vs song ads (Sep 2026))
  5. Label realistic AI people and voices, and never script an AI person as a real customer who 'used it for two weeks'. The FTC's final rule (in force since 21 Oct 2024) bans fake testimonials, including AI-generated ones, with civil penalties. TikTok has required an AI disclosure on ads with realistic AI people, voices or product imagery since 21 July 2026, with penalties up to a ban (agency summary). (FTC, final rule banning fake reviews and testimonials)

Stopped working:

Prompts that work

Templates from the creators of the best current results and from the model guides. Fill in the {CAPS}.

Phone-real creator still (JSON spec) (seedance_2_5)

Casts the creator before any ad clip: run it as a 4 s, 9:16, 1080p Seedance 2.5 take, pause on the best frame and save it as Image 1 for every shot. Adapted from DavidRunsAds: the exact JSON behind a Seedance 2.5 character (1.4k bookmarks).

{
 "output": {"aspect_ratio": "9:16", "style": "photorealistic candid front-camera selfie, casual social-media look; not studio, not illustration, not 3D render", "summary": "{ONE SENTENCE: who, where, time of day, what they are doing}"},
 "subject": {"count": "exactly one person", "apparent_age": "{AGE_RANGE}", "build": "{BUILD}", "identity": "concrete features, not a type: {FACE_SHAPE}, {EYES}, {BROWS}, {NOSE}, {LIPS}", "expression": "{EXPRESSION}, looking into the lens"},
 "skin": "real skin: visible pores, mild redness, uneven tone, {FRECKLES_OR_MARKS}; not airbrushed",
 "hair": "{HAIR_COLOR_AND_STYLE}, individual strands, flyaways at the crown",
 "clothing": "{GARMENT}, {FABRIC}, natural folds, no logos or lettering",
 "environment": "{LIVED_IN_ROOM} with 3-4 named objects, 1-3 m behind, slightly soft",
 "lighting": "{WINDOW_DAYLIGHT_OR_LAMP_PLUS_SCREEN_GLOW}, soft, low contrast, no rim light",
 "camera": "handheld phone front camera at arm's length, ~24-26 mm, small-sensor noise, mild edge stretch, no bokeh",
 "composition": "chest-up, face upper-middle, camera rolled a few degrees",
 "color_grading": "muted, natural white balance, no teal-orange grade, no heavy HDR",
 "negative_constraints": ["no text, UI, captions or watermarks", "no extra people", "no visible phone", "no invented accessories", "no plastic or waxy skin, no perfect symmetry", "no extra or fused fingers", "no studio light or hard flash"]
}

Seedance 2.5 reference-led UGC take (seedance_2_5)

The ad clip itself: your creator (Image 1) and product (Image 2), moved like a real TikTok you picked (Video 1), with native voice; one 15 s clip per prompt. Adapted from BytePlus: Dreamina Seedance 2.5 prompt guide (asset mapping, timeline, dialogue format).

Raw vertical 9:16 phone video, one UGC take for {PRODUCT}. Image 1 is the creator: keep her face, hair and outfit exactly. Image 2 is the product: keep its shape, colour and label exactly. Strictly refer to Video 1 for the handheld camera movement, framing, pacing and how a real creator behaves on camera; do not copy its person, words or on-screen text.
Setting: {LIVED_IN_ROOM}, {LIGHT}. Phone front-camera quality, slight handheld sway, autofocus breathing.
0-2s: Opens mid-sentence, the product from Image 2 already in her hand near the lens. Creator's line (casual, a little fast): "{HOOK_LINE}"
2-6s: She {DEMO_ACTION_1} close to the lens, label readable. Creator's line: "{PROBLEM_LINE}"
6-11s: Close-up of {DEMO_ACTION_2} showing {VISIBLE_PROOF}. Creator's line: "{PROOF_LINE}"
11-15s: She looks back into the lens, product beside her face, small laugh. Creator's line (calm): "{SOFT_CTA}"
Audio: only her voice, room tone and handling sounds. No BGM.
[Constraints] One creator, same face and outfit throughout; five fingers per hand; the product never changes shape or duplicates. No subtitles, captions, extra logos or watermarks.

Seedance one-take street interview (seedance_2_5)

A 15 s vox pop: off-screen interviewer, fuzzy mic in frame, the passer-by pulls the product out at about 10 s. Adapted from Zyvex_0x: full Seedance 2.5 street-interview prompt (rules, camera, defect list, timed storyboard).

Native 15-second vertical 9:16 street-interview clip: one continuous handheld take, no cuts, on a {CITY_STREET} in {TIME_OF_DAY}. It must look like a real viral man-on-the-street clip, not an ad. Image 1 is the passer-by: keep face, hair and outfit exactly. Refer to Video 1 only for camera movement, street look and how people react to a mic.
CORE RULES: one adult in focus. The interviewer is never seen except his RIGHT hand holding a fuzzy grey mic from the lower-left; his voice is off-screen. Five fingers per hand. No readable text or logos except on {PRODUCT}.
VOICE: {VOICE: age, accent, pace}. Interviewer upbeat, close to the mic. Everything through one handheld mic: street ambience, wind rustle.
CAMERA: older consumer camera, operator 1.2 m away at eye level, waist-up, natural shake, small reframes, brief autofocus hunt; never gimbal-smooth.
LOOK: soft, slightly muted colour, imperfect white balance, real skin texture, {OVERCAST_DAYLIGHT_OR_STREETLIGHT}; no beauty filter, no grade.
STORYBOARD:
00.0-02.5 INTERVIEWER: "{QUESTION}" She smirks, eyes up.
02.5-07.0 Passer-by's line (honest, unhurried): "{ANSWER_WITH_PAIN}"
07.0-10.0 INTERVIEWER: "{FOLLOW_UP}" She reaches into her bag.
10.0-13.0 She holds {PRODUCT} beside her face, label to the lens: "{ONE_LINE_REASON}"
13.0-15.0 She shrugs and laughs; the mic hangs; the camera holds.
SOUND: city ambience, distant traffic, no music. No subtitles, no watermark.

Trending formats to remake

To remake one: download the example with the media-download skill, take a frame a second and the transcript, write a beat list (time, shot, camera, on-screen text, sound), keep what the format needs, swap in the person's product or character (never the creator's face, voice, logo or footage), show the beat list and stills before spending, then make it with the skill named below.

  1. Unboxing or in-hand demo creator clip: A creator, real or AI, opens or uses the product on camera. The product is visible by second 2, with one-word captions and no talking-head preamble. 15-30 s. Why it spreads: Unboxing (9.8%) and demo (8.1%) are the two highest hit-rate visual formats in Motion's 2026 benchmark of 578,750 Meta creatives, ahead of testimonials (6.5%). Keep: Keep the product already in hand (or the box already open) by second 2, one-word captions, no talking-head preamble and your real product photo as the reference; swap the product, the room and the creator. Make it with fal UGC (fal-ugc).
    • Example to remake: @frankyecom on X, 2026-04-22, 286K views, 1K saves: 23 s AI creator unboxing a duvet: a stack of branded boxes in frame 1, product out of the bag by second 3, one-word captions, no one talking to camera. Built with his 'v5.1 realism' workflow; no prompt shared.
  2. Micro-drama ad (a show, not a pitch): A 1-3 minute scripted scene with recurring characters and conflict. The product appears late and solves the problem. A variant remakes an old TV format, such as a shopping-channel live demo. Why it spreads: Franky Shaw's drama-ad post drew 1.29M views and 6,224 bookmarks, and several ad builders (Max Battle, 0x ROAS, Miko) describe drama ads as the format 'ripping' on Facebook in Q3-Q4 2026. People skip ads, not stories. Keep: Keep a conflict in the first 3 s, recurring characters with locked faces, at least half the runtime as story before the product, and the product as the turn that fixes it; swap the setting, the conflict and the cast. Make it with Consistent Characters (varg-consistent-characters).
  3. Podcast-clip ad: A 30-70 s clip that looks cut from a podcast: one studio set, mic in frame, a host and a guest (or one guest), a captioned hot take, and the product as the guest's recommendation. Why it spreads: Operators call it the most native format: 'if I could only keep 1 ad format, it'd be AI podcast ads' (a brand owner claiming $6M a month). One script becomes many expert personas, and viewers drop their guard because it looks like content. Keep: Keep one studio set with the mic in frame, a captioned hot take as the first line, the guest as an expert persona and the product only as their recommendation near the end; swap the topic, the guest and the product. Make it with Turn a script into a two-person podcast video (varg-podcast-video).
    • Example to remake: @Mho_23 on X, 2026-08-16, 25K views, 549 saves: 68 s single-guest podcast-set talking head cut with archival photos and engraved illustrations; the product (a cacao-and-dates bar) is held up only at the end. Author says it was fully automated for $10.26; no prompt shared.
  4. AI street interview (vox pop): A handheld street interview: a mic pushes into frame, a passer-by answers one question, then pulls the product from their bag. 15-45 s, question as the first caption. Why it spreads: This Seedance 2.5 vox pop drew 436 bookmarks on 18k views (a 2.4% save rate). Ad builders list street interviews next to podcasts and dramas as the formats AI now makes cheaply. Keep: Keep one handheld take, the interviewer's hand and fuzzy mic in frame, the question as the first caption and the product pulled from the bag at about 10 s; swap the city, the question, the passer-by and the product. No dedicated skill yet: build it with varg-ai.
    • Example to remake: @DavidRunsAds on X, 2026-10-21, 18K views, 436 saves: 15 s vox pop on a city street made with Seedance 2.5: the interviewer's mic and hand stay in frame, the question is the first caption, she pulls the product out at about 10 s. It uses a real brand (The Ordinary); swap in your own. No prompt shared.

Open-source repos

Set up varg

varg is one API key for the models these skills drive — Seedance 2.5, Kling 4, Veo 3.1, Wan 3, Nano Banana 2.1, GPT Image 2.5, ElevenLabs v4 and 200+ more — billed in credits, with a cost estimate before anything runs.

Related tutorials