nim-ugc-prompts

name	nim-ugc-prompts
description	Use when writing, improving, planning, or generating Nim/Seedance UGC video prompts for TikTok, Reels, Shorts, creator ads, direct-to-camera hooks, selfie videos, iPhone footage, testimonials, street interviews, mirror try-ons, outfit changes, product/location walkthroughs, VO, lip sync, audio references, or image-to-video prompts with character/location/product references.

Nim UGC Prompts

Overview

Create copy-ready English UGC video prompts for Nim image-to-video and text-to-video workflows. The output should feel filmed by a real person on a phone: imperfect framing, ordinary timing, natural delivery, realistic skin/fabric/room behavior, and clear reference roles.

For detailed examples, load:

• references/ugc_prompt_patterns.md for distilled patterns.

Nim MCP Rules

When actually generating through Nim:

1. Discover the live video model with models_explore; prefer the current Seedance image-to-video model when references drive the result.
2. Inspect the exact generationContract with models_explore action: "get".
3. Upload every image/audio reference with media_upload; use only returned Nim file_url values in fileInputs.
4. Pass only parameters supported by the selected contract: prompt, requestedAspectRatio, mediaLength, resolution, fps, keepSound, seed, batchSize, or any exact contract fields.
5. Poll get_generation_status until a terminal status returns a real media URL.

Default targets:

• Format: vertical 9:16 unless the user asks otherwise.
• Duration: 6-10 seconds for single UGC beats; 10-16 seconds for multi-cut try-on, walkthrough, or story sequences.
• Reference stills: use Nano Banana Pro 2K for quick identity/product/location refs, or GPT Image 2 medium 2K when controlled polished still anchors are more important.

Intake

Extract or assume:

• platform and duration;
• product/offer/topic;
• creator identity, audience, and emotional arc;
• references and their roles;
• whether the clip is single-take, multi-cut, selfie/POV, mirror, static tripod, or location walkthrough;
• VO, dialogue, native audio, lip sync, and whether subtitles/text overlays are allowed.

Ask only when a required reference or audio asset is missing.

Reference Roles

Write reference roles at the top of the prompt before scene directions:

REFERENCES:
@image1 - main location, preserve layout exactly.
@image2 - main character identity, face, hair, body proportions exactly.
@image3 - outfit/product/object reference, preserve design exactly.
@audio1 - voice, tone, cadence, pronunciation, pacing and vocal texture exactly.

Use stable role names:

• main character;
• additional character/friend/interviewer;
• outfit/body proportions;
• product/object;
• main location/exterior/interior;
• phone/mirror/tripod/camera reference;
• audio/voice/song/live performance.

Never let references merge. If @image2 is outfit only, say not to use it for face or hairstyle.

Prompt Skeleton

Use this structure for most UGC prompts:

REFERENCES:
@image1 - ...
@image2 - ...

AUDIO / VO:
...

IMPORTANT:
...

STYLE:
UGC mixed with documentary realism, social media ad, TikTok/Reels/Shorts, authentic creator content, smartphone footage, natural imperfections, no polish.

FORMAT:
Vertical 9:16. <duration> seconds. <front camera / rear camera / mirror / tripod / handheld>.

CAMERA:
...

CHARACTER:
...

LOCATION:
...

SCENE / SHOTS:
0:00-0:03 ...
HARD CUT.
0:03-0:06 ...

NEGATIVE:
No subtitles unless requested. No text overlays. No beauty filter. No skin smoothing. No CGI. No morphing. No flickering. No ghosting. No hidden transitions.

UGC Prompt Patterns

Choose one primary pattern:

Pattern	Use for	Core camera behavior
Direct-to-camera hook	creator ad, testimonial, confession	front camera or handheld medium shot, natural eye contact
Multi-cut hook	punchier 6-10s social ad	hard cuts only, same location/character continuity
Selfie walkthrough	store, travel, event, venue	front-camera selfie then rear-camera POV
Mirror try-on	fashion/outfit, fitting room	mirror geometry, phone visible if needed, body/outfit clear
Static tripod/home	performance, demo, try-on sequence	locked framing, background unchanged
Street interview	social proof, comedy, dating, vox pop	handheld interviewer camera, spontaneous timing
Concert/event UGC	music, crowd, live reaction	selfie camera, crowd density, live audio and changing exposure

Load references/ugc_prompt_patterns.md for a richer pattern library.

Writing Rules

• Make ordinary behavior specific: blinking, breathing, posture shifts, small pauses, glances away, half-smiles, subtle laughs, natural hand gestures.
• Use smartphone image behavior: natural HDR, autofocus breathing, auto-exposure adaptation, tiny wrist shake, real sensor noise, practical room light.
• Preserve the reference environment exactly when the prompt says the scene happens inside one location.
• Use HARD CUT for edits. Do not use hidden transitions, morphing, whip-pan disguises, or magical outfit swaps unless requested.
• For lip sync, specify exact spoken lines and performance timing. Do not add subtitles unless requested.
• For static tripod prompts, explicitly forbid camera shake, zoom, panning, reframing, and background movement.
• For handheld prompts, specify which hand holds the phone and what the other hand can do when it matters.
• For mirror prompts, explicitly describe reflection logic and avoid impossible camera/mirror geometry.
• For try-on or product sequences, lock camera position, background, exposure, scale, and reference role across every cut.

Output Format

Return this by default:

## Reference Plan
- @image1: ...

## Final UGC Video Prompt
[copy-ready English prompt]

## Nim Generation Notes
- Discover model target: Seedance image-to-video / text-to-video as required
- Aspect ratio: 9:16
- Duration: ...
- Required fileInputs: ...

## Negative Prompt
...

If the user asks for JSON, return:

{
  "reference_plan": [],
  "prompt": "...",
  "nim_generation_notes": {
    "discover_model_target": "...",
    "aspect_ratio": "9:16",
    "duration_sec": 0,
    "required_fileInputs": []
  },
  "negative_prompt": "..."
}

Quality Bar

• The clip reads as a real phone recording, not a polished commercial.
• Reference roles are unambiguous and do not merge.
• Face, outfit, product, and location locks survive across cuts.
• Delivery sounds conversational, not scripted.
• Camera behavior matches the declared capture mode.
• Timing is feasible for the requested duration.
• Negative constraints block common AI failures without fighting the brief.