USE THIS SKILL whenever the user mentions UGC, a UGC video, a UGC 视频/脚本, a short video / short-form video / 短视频, or any TikTok / Reels / Shorts / 抖音 / 快手 / 小红书 clip — and whenever they want a "get ready with me" / GRWM / try-on / haul / unboxing / 开箱 / product review / 测评 / 种草 / 带货 / 口播 / influencer / 达人 / 博主 video. ALSO trigger when the user names any AI video model — Seedance / 即梦 / Kling / 可灵 / Veo / Sora / Hailuo / 海螺 / Doubao / Volcengine — and wants a prompt or a video from it, OR just uploads a product / person photo and says "拍成视频 / 做成短视频 / make a video / run it / 跑一条 / 出个视频". Trigger even if they only describe a product, a creator, or a scene and never literally say "UGC". This is the right skill for ANY request to write a video-model prompt or to actually render a short UGC-style clip. What it does: writes production-ready prompts for AI video models (Seedance, Kling, Veo, Sora, Hailuo) that generate authentic, viral-feeling, handheld-iPhone UGC short videos. It first runs a scene-coherence pre-check (product / location / persona / reason-to-film) so the clip reads like a real person filmed it, not a glossy ad. Outputs a clean, copy-pasteable prompt (画面 visuals + 台词 dialogue + camera + audio) tuned to the exact model named, keeping @-style asset placeholders (@avatar, @product, @image_1). It also bundles tools that call the vendors' official APIs to generate the actual mp4 end-to-end (submit prompt + reference images, poll, download) — including a verified Kling Omni dual-image path that locks BOTH a person photo and a product photo to their originals at once (the API equivalent of 即梦's @图1 @图2).

UGC Video Prompt Generator

What this does and why it works

This skill writes the text prompt that goes into an AI video model. The goal is a video that looks like a real person filmed it on their phone — not a polished ad. That "authentic amateur" quality is what makes UGC spread, and it comes almost entirely from how the prompt is written: the model will happily produce glossy, over-graded, tripod-stable footage unless you actively tell it not to.

So the whole craft here is directing the model toward realness (handheld micro-shake, natural daylight, real skin, no color grading, casual self-aware dialogue, a tiny imperfection) while keeping the product or character as the clear star of the shot.

You are writing for a specific model each time. Models differ a lot on the two things that matter most — whether they generate spoken dialogue/audio and how they take in a reference image of the product/person. Get the target model first, then write to its strengths. Details per model live in references/ — read the relevant one before finalizing.

Step 1 — Get just enough to start

UGC prompts need very little to get going. Ask only what you genuinely can't infer, and ask it in one batch:

What's the video about? (a product to feature, or a content concept — e.g. "GRWM with a chaotic outfit", "unbox these sneakers", "review this serum")
Which model? (Seedance / Kling 可灵 / Veo / Sora / Hailuo 海螺 — this changes dialogue syntax, audio, params, and how assets are referenced). If they don't know or don't care, default to a model with native audio since UGC lives on spoken hooks — pick Veo 3 and say so.

Nice-to-have, infer if not given: who's on camera (creator persona), language of the spoken lines, vibe (playful / deadpan-cool / genuine-excited), and whether there's an avatar/product asset to reference.

Don't over-interview. If the user gives you a product and a model, you have enough — write a strong default and let them react to it. Iterating on a real draft is faster than answering ten questions.

Step 1.5 — Context coherence pre-check (do this before writing)

This is the highest-leverage step in the whole skill. The "this is an ad / this is AI" feeling comes far more from a scene that doesn't cohere than from any camera setting. Real UGC is believable: a specific person has a plausible reason to be filming this thing, in this place, in this way. Before writing any prompt, lock these four and confirm they cohere:

PRODUCT — what is it, and where does it physically live? (a desk ornament lives on a shelf; a serum lives in a bathroom; sneakers live by the door.)
LOCATION — an ordinary, specific, lived-in place that is a plausible home for THIS product. Default to the creator's own space (bedroom, car, kitchen, desk, bathroom mirror). Only leave home for an aspirational setting (yacht, villa, hotel, pool) if the brief explicitly asks for it.
PERSONA — who is filming, and what's one concrete, persona-true detail of their real space?
REASON-TO-FILM — why is the camera rolling right now? (it just arrived / I'm packing / a friend asked / I caught myself using it.)

The one-sentence test: "Why is THIS person, with THIS product, in THIS place, filming right now?" If you can't answer it in one believable sentence, the scene will read as staged. Fix the location or the reason — not the lighting.

Plausibility of the rig: a one-hand selfie creator can shoot at arm's length in an ordinary room; they cannot simultaneously frame a wide aspirational backdrop AND a tight product macro — that needs a crew the persona doesn't have. If the concept wants both, drop one. (Never write in a second person unless the beat needs one — models render an extra subject and break the selfie intimacy.)

Register match: keep location lavishness, packaging tier, and speaking tone on ONE register. Confessional best-friend tone ("sisters, you have to see this") → casual setting + casual reveal ("treated myself" / "this was a gift") — not a yacht or a pristine gift-box hero shot. A premium product is most believable as "I treated myself," filmed at home.

If the user explicitly asked for an aspirational/messy/specific setting, honor it — this gate governs the silent default, not the user's stated wish.

Step 2 — Pick the structure

There are two proven structures in this genre. Choose based on the content, and say one line about why.

Script form (no timecodes) — best for a single continuous moment with a beat or two: a GRWM with a friend interrupting, a quick product reaction, a genuine testimonial. Reads like stage direction. This is the default for most clips, especially on models that don't do reliable multi-shot.

Style: [stacked vibe tags — UGC, get ready with me, iPhone front camera, playful energy]

[Scene/room description — lived-in, not styled: an ordinary everyday place that is a plausible home for this product, with a few real details that look used rather than placed, kept to the background so the subject/product stays the clear star]

[Camera spec as its own line — shot on iPhone front camera, vertical 9:16, slight handheld movement, real skin tones, no color grading]

[Action + dialogue, alternating. Dialogue in the model's preferred syntax.]
[A small beat or interruption — the thing that makes it feel real and watchable.]
[A closing action — step back to show full outfit / final pose / clip cuts mid-motion.]

[One-line vibe summary — natural messy UGC vibe, confident energy, light humor]

Timecode form — best for try-on hauls and multi-stage sequences where the outfit/look changes via jump cuts. Each beat gets a timestamp.

A [N]-second vertical (9:16) UGC [type] video filmed on a smartphone. [Subject + setting + camera feel in one sentence.]
0–3s: [beat — what they wear/do + expression]
3–5s: [jump cut — next stage]
5–8s: [jump cut — next stage]
...
[final timecode]: [final pose, holds a beat, clip cuts]
Style: [aesthetic summary — quick jump cuts, handheld shake, natural light, the product is the star]

Only use timecodes if the model supports multi-shot well (Kling 3.0, Sora 2, Seedance) — otherwise the model may ignore them and you've added noise. When in doubt, script form. The references/ file for the model tells you.

Step 3 — Write the prompt

Apply these patterns regardless of structure. They're what separate a UGC prompt from a generic video prompt.

Real-feel anchors (the most important part)

Models default to polished. You must explicitly request the opposite. Pull from:

Location (most important — set the scene before you light it): authentic UGC is filmed in ordinary, specific, lived-in places — a real bedroom, a car, a cluttered kitchen counter, an office desk, a bathroom mirror — not aspirational showrooms. DEFAULT to an unglamorous everyday place that is a plausible home for THIS product. Name a concrete mundane place ("her own messy bedroom", "the driver's seat of her car"), never a mood word ("aesthetic / minimalist / bright / vacation feel") — mood words render as stock B-roll. Use an aspirational/branded setting only if the brief asks for it, and even then keep it handheld, candid, un-staged.
Camera: "shot on iPhone front camera", "handheld selfie perspective", "subtle micro-shake", "slight handheld movement", "natural smartphone-lens look"
Composition (optional): slightly off-center and casual — subject a bit to one side rather than dead-center, product entering at a natural angle rather than squared to the lens. Not a balanced commercial product-hero shot. (When she holds up the product, it still stays the clear, in-frame subject.)
Color/grade: "no color grading", "no cinematic grading", "real skin tones", "slightly warm tones", "no filters". For model-generated faces (text-to-video), add "natural skin texture — visible pores and fine lines, no beauty smoothing, no poreless glow" plus avoid-clause "no beauty filter, no skin smoothing" — "real skin tones" alone only fixes hue, not the waxy AI face. (Skip the texture cue on image-to-video — the face comes from the uploaded photo; don't override it.) Do NOT request "natural HDR" — HDR's whole job is to remove the highlight clipping that signals real capture, so it nudges toward the polished look you're fighting.
Light: "soft natural daylight from a window", "no ring light" (this last one is gold — it kills the tell-tale AI/influencer over-lighting)
Environment — lived-in, not styled: name 2–3 ordinary background details that look used, not placed — kept at the frame edge and soft/out of focus, while the product/subject stays the sharp, centered star. Use neutral, persona-true objects (a charging cable trailing off the nightstand, a half-drunk mug, a couple of stacked books with one askew, an unmade-bed corner, a hoodie over a chair). Drop the singular "one" and the word "deliberate" — a single tidy prop reads as art direction. Keep OUT the set-dresser tropes ("folded towel / small plant / simple ceramics") and anything gross (used tissue, laundry pile — degrades product appeal and renders ugly). Vary the objects across prompts — a fixed list becomes its own tell. Clutter is background texture only — never centered, never dense enough to compete with the product. If in doubt, less.

Don't dump all of these — pick 4–6 that fit. Too many and the model gets confused; too few and it reverts to glossy. If you're over budget, drop redundant CAMERA anchors first (keep one of handheld/micro-shake), but ALWAYS keep the lighting + grade-suppression anchors (soft window daylight, no ring light, no color grading, real skin tones) — those are the load-bearing anti-gloss levers.

Dialogue — write it natural, and check the model's syntax

Real UGC speech is casual, self-interrupting, a little messy:

Use contractions, filler, trailing off: "Okay, I'm getting ready and I don't know if this outfit is crazy or—"
Break the fourth wall: "Anyway… I kinda love it.", "You are welcome."
Keep each line short (clean lip-sync needs ≤ ~8s of speech per beat).

Vary the energy; don't sustain it. Constant high enthusiasm across every beat reads as an ad — the giveaway isn't excitement, it's that every line is photogenic delight on-message. Give the clip a flat, offhand baseline (like talking to one friend) and let ONE moment carry a genuine reaction. A dry aside ("…okay that's actually kind of nice") beats four enthusiastic lines. Low-key/deadpan is valid and underused — don't default to "genuine-excited."

Gaze + behavior (a little humanness goes a long way):

Don't hold a single locked smile + dead-on lens stare for the whole clip. Let attention land on the product for part of it and meet the lens only briefly ("mostly looks at the product, glances up to the lens once while talking, eyes fall back"). (No blink instructions on 4–8s clips — they render as darting eyes.)
Optionally add one human-friction beat: a tiny false start ("wait— okay"), pushing hair off her face, a quick "is this even recording?" glance, a small self-conscious laugh. Don't fumble/nearly-drop the product, and don't have her reframe or check the phone mid-take (that's a second camera move). If you use a behavioral beat, you can drop the environmental one — don't stack both plus a camera move into a <8s clip.

Critical: dialogue syntax is model-specific. Veo uses Character says: and quotes can trigger unwanted on-screen captions; Sora uses a labeled Dialogue: block; Hailuo and Seedance 1.0 produce no audio at all (write the lines as intent, plan to dub separately). Always read references/<model>.md and use that model's exact convention before finalizing.

The hook and the beat

Viral UGC earns the first 2 seconds. Open on a hook:

Visual: open already in motion — she's mid-sentence, product already in hand; or bring the product up close toward the lens at a natural off-center angle, the room still visible behind it (close and prominent, not gallery-centered, never clipped). Avoid the choreographed "lean in fast + wide eyes" lunge — it reads as performed.
Verbal: "okay wait—", "There are TOYS in the sole.", a confident claim. Convey energy through expression and a verbal cold-open, not speed.

Avoid speed words on the SUBJECT (fast, quickly, lurch, lunge). On Seedance the word "fast" is the single biggest documented quality degrader, and a fast move toward the lens produces face-warp / rubber-arm morph on Kling and Hailuo too. (Camera-move terms like Kling's "whip-pan" are fine where the model's reference lists them — this ban is on subject speed, not camera vocabulary.)

Then give it one beat — a small reversal or surprise that makes it feel unscripted and re-watchable: a friend wandering into frame and getting shooed out, a product detail revealed ("a little bear in there"), a playful contradiction ("it's a little chaotic… but it works"). Keep the beat consistent with the premise: if she already owns and loves the product, she can't "just now discover" a basic feature — make the reveal about the VIEWER ("you can't see this in the listing photos"), not a fake first-time reaction. Fold an implicit reason-to-film into the opening (it just arrived / I'm packing / a friend asked).

For a single continuous handheld/selfie take, use one primary camera move — don't chain focus or framing moves (face → product → macro) in one prose line; on single-shot models that renders as a floaty continuous drift or gets ignored. Let the subject's motion reveal the detail instead. (Multi-shot models — Sora 2, Kling 3.0, Seedance 2.0 — can use labeled beats; see Step 2.)

Make the product/person the star

For product videos: give the product real screen time and describe it concretely (materials, colors, the one distinctive detail) so the model renders that product, not a generic stand-in. But show it being physically handled, not posing for a commercial:

A visible brand logo or a pristine gift box centered in frame is the single strongest "this is a paid ad" tell. Prefer to show the product as something already owned and used (slightly handled, out of its box). Primary fix: omit the standalone branded box from the frame, and add no logos, no packaging, no gift box to the avoid/negative clause. If packaging must appear, keep it to one incidental detail off to the side, partly out of frame — never a second hero object competing with the product. (i2v note: if a pristine box is supplied as a reference image, prompt text won't make the model "use" it — just don't feed a pristine-box photo.)
Don't write "rotated slowly to catch light" — that trio (slowly + rotate + centered) is the motorized-turntable recipe. Instead add exactly ONE grip/weight cue: she turns it in her hand, pausing when the light catches the [hero detail], then shifts her grip — the highlight slides and briefly flares as her hand moves. This breaks the AI turntable look. Never occlude the hero detail (no thumb over the pattern), and never stack re-grip + dip + fumble in one short beat (warps fingers).

Closing

Pick ONE closing mode — don't staple two together. For amateur realness prefer the motion cutoff: the camera is still moving and slightly off-target at the cut (arm starting to lower, frame tilting away), product still roughly in shot — not a held, perfectly-composed pose. The alternative is a clean final pose held for a beat with a satisfied micro-smile. Do not combine "ends mid-motion" with "holds a satisfied smile / final pose held" — that resolves toward a stabilized ad ending, which is the opposite of what you want. Keep the product in frame through the cut.

Step 4 — Asset placeholders

Keep @-style placeholders so the user can wire in their own assets. Use semantic names, not invented IDs:

@avatar — the creator/person on camera
@product — the featured product
@image_1, @image_2 — specific reference images (e.g. packaging, a logo)

Place them inline where the asset is referenced, e.g. "@avatar holds up @product to the front camera" or "first opens the box @image_1 then takes @product out". Tell the user in a short note that they should replace these with their platform's real asset IDs, and that how the asset is actually bound depends on the model (most models take the product/person as a separate uploaded reference image — first frame or subject reference — not literally as an in-prompt token). The per-model reference file explains the real binding for that model.

Step 5 — Deliver

Output the finished prompt in a clean code block so it's one-click copyable. Then, briefly (a few lines, not a wall of text):

Note the model it's written for and any params to set (aspect ratio 9:16, duration, audio on/off, negative prompt for captions if relevant).
Note what to do with the @placeholders.
If the model can't do audio (Hailuo, Seedance 1.0), flag that the dialogue needs separate dubbing/lip-sync.

If the user named no model, write for Veo 3 by default (native audio suits UGC's spoken hooks), output the prompt, and tell them you can retarget it to Kling/Sora/Seedance/Hailuo if they prefer.

Step 6 — Optionally generate the actual video

The prompt is the deliverable, but the skill can also turn it into a real mp4 by calling the vendor's official API. Offer this whenever the user seems to want the finished video (they uploaded assets, said "make the video", or asked to "run it"). There are two paths — pick by how the assets must be bound.

6a. Kling Omni — the dual-image path (RECOMMENDED when there's a person AND a product)

This is the one path that locks both a person image and a product image to their originals at once — the true API equivalent of 即梦's @图1 @图2. It's a bundled, ready-to-run Node toolkit (scripts/kling/, zero npm deps, Node 18+), verified working end-to-end. Use it as the default when the user has two real assets (avatar + product) and wants them both faithful.

# 1. credentials — written once to ~/.config/kling/.credentials (INI):
#    [default]
#    access_key_id = <AK>
#    secret_access_key = <SK>
#    (or set KLING_TOKEN for a session). Verify with:
node scripts/kling/kling.mjs account --costs

# 2. generate (two --image inputs, comma-separated → auto-routes to omni-video):
KLING_MEDIA_ROOTS="<dir with the images>" \
node scripts/kling/kling.mjs video \
  --model kling-v3-omni \
  --prompt "<<<image_1>>>中的人物 拿着 <<<image_2>>> 的产品，手持自拍展示…" \
  --image "person.png,product.jpg" \
  --aspect_ratio 9:16 --duration 5 --mode pro --sound on \
  --output_dir "<output dir>"

Critical details (all learned the hard way — honor them):

Endpoint is api-beijing.klingai.com (CN) / api-singapore.klingai.com (global) — NOT api.klingai.com. The toolkit auto-probes the right one. The bare api.klingai.com returns a misleading code=1002 Auth failed even with a valid key. (code=1000 = bad secret; code=1002 = account/endpoint mismatch.)
Reference images by index in the prompt with <<<image_1>>> / <<<image_2>>> (and <<<element_1>>> for a registered subject). Images map to indices by their order in the --image list. This is Kling's @图N.
Multiple --image (comma-separated) auto-routes to the omni-video endpoint and builds image_list. One image stays on plain image2video.
Real people are allowed here. Kling omni accepts a real-person photo + a product photo together — unlike Seedance 2.0, which rejects real faces (InputImageSensitiveContentDetected). This is the main reason to prefer Kling omni for person+product UGC.
Local image paths need an allow-root: set KLING_MEDIA_ROOTS=<dir> (comma- separated) or KLING_ALLOW_ABSOLUTE_PATHS=1; otherwise only the cwd is readable. URLs always work.
--model kling-v3-omni (default) or kling-video-o1 (o1 has no --sound). --mode pro|std, --duration 3–15s, --sound on|off.

6b. generate.py — the single-vendor path (Seedance / Veo / Sora / Hailuo, or single-image Kling)

scripts/generate.py is a Python CLI over all five vendors for the single reference image case (first frame / subject reference). Use it when the user named a specific non-Kling model, or only has one asset to lock.

pip install -r scripts/requirements.txt
# Veo, vertical, with caption suppression:
python scripts/generate.py --model veo \
  --prompt-file prompt.txt --out video.mp4 \
  --aspect 9:16 --duration 8 --negative "no subtitles, no text, no captions"
# Seedance 1.0-pro accepts a real-person first frame (2.0 rejects real faces):
python scripts/generate.py --model seedance --model-id doubao-seedance-1-0-pro-250528 \
  --prompt-file prompt.txt --image avatar.png --out clip.mp4 --aspect 9:16

Key things to get right:

Save the prompt to a file first (prompt.txt), pass --prompt-file — UGC prompts have quotes/em-dashes/newlines that break inline.
Bind assets via --image / --image-tail, not the @placeholders. Accepts a local path or http(s) URL — except Veo and Sora, which need a local file.
Match flags to the model: --audio for Seedance 2.0; --negative "no subtitles..." for Veo; --size for Sora. Defaults in scripts/README.md.
Seedance real-person rule: 2.0 rejects real faces → use --model-id doubao-seedance-1-0-pro-250528 for a real-person first frame (no audio, dub separately). 2.0 is fine for product-only first frames (with --audio).
Tell the user which env vars to set (table in scripts/README.md) and which region/base URL matches their key.

If a model can't do audio (Hailuo, Seedance 1.0), remind the user the spoken lines need separate dubbing/lip-sync. Read scripts/README.md for the env-var table, per-model examples, and URL-expiry gotchas.

Model reference files

Read the relevant one before finalizing — they carry the exact syntax that makes or breaks the output:

references/seedance.md — Seedance 1.0 (silent, first/last frame) vs 2.0 (@Image1 numbered refs + native audio); -- params; 9:16, duration
references/kling.md — 可灵 formula (镜头+光影+主体+运动+场景+氛围); Omni dual-image binding (<<<image_1>>>/<<<image_2>>>, the working scripts/kling/ toolkit, api-beijing.klingai.com endpoint, real-person allowed); Kling 3.0 Omni audio
references/veo.md — Character says: dialogue, the quotes→captions trap and no subtitles, no text, no captions fix; ingredients/reference images; 4/6/8s
references/sora.md — labeled Cinematography: / Actions: / Dialogue: / Background Sound: template; characters API; 9:16 sizes, durations
references/hailuo.md — silent (dub separately); [Pan left]-style bracket camera commands; S2V single-image subject reference; 6/10s

Quick reference: model capabilities

Model	Native audio/dialogue	Asset reference	Multi-shot/timecodes	9:16	Typical duration
Veo 3	Yes (T2V)	up to 3 ref images	No (one shot)	Yes	4/6/8s
Sora 2	Yes	input_reference / Characters	Yes (labeled beats)	Yes	4/8/12/16/20s
Seedance 2.0	Yes	`@Image1`…`@Image9`	Yes (prose)	Yes	4–15s
Seedance 1.0	No	first/last frame	No	Yes	2–12s
Kling 3.0 Omni	Yes	dual-image: person+product, real faces OK (`<<<image_N>>>`)	Yes (≤6 shots)	Yes	5/10s
Hailuo	No	S2V single image	No	Yes	6/10s

name

ugc-video-prompt

description