nim-hookgen-seedance

name

nim-hookgen-seedance

description

Author a production-ready Seedance 2 hook video for short-form UGC (TikTok / Reels / Shorts) and generate it through the Nim MCP. Use when the user wants a talking-head "hook" ad/UGC clip from a product brief, asks to "make a hook video", "generate a UGC hook", "spoken hook ad", or "Seedance hook". The agent acts as the prompt-writing LLM: it intakes the brief, proposes 5 spoken hook options, collects character/product/location reference images, assembles the Seedance brief, rewrites it into a final Seedance prompt, and generates the video on Nim Seedance 2. NOT for: long scripts, code, text-only answers, or audio-only tasks.

HookGen Seedance

Turn a product brief into one spoken-hook UGC video. This skill writes the spoken hooks and rewrites the assembled brief into the final Seedance prompt.

UX rules

1. Be concise. Don't dump raw IDs, JSON, briefs, or tool plumbing into chat. Report the finished result, not the mechanics.
2. No jargon. Don't narrate "calling models_explore" or "polling status".
3. Reply in the user's language. Detect it from their message; keep technical values (9:16, 720p, @Image1) and the chosen spoken line as-is.
4. One thing at a time. Ask only for a genuinely missing required input. Pick sane defaults and move.
5. Never invent the result link. generate_video returns no link; only the URL from get_generation_status is real.

Fixed defaults

These mirror the verbal-hook path of the pipeline. Do not change unless the user asks.

Setting	Value	Source
Hook video duration	6 seconds	HOOK_VIDEO_DURATION
Aspect ratio	9:16	DEFAULT_ASPECT_RATIO
Resolution	720p	DEFAULT_RESOLUTION
Model	Seedance 2 (reference-to-video when images exist, else text-to-video)	FAL_*_MODEL
Audio	generated	generate_audio: true
Modes	Paid Ads or UGC	MODES
Verticals	Beauty, Food & Drinks, Fashion, B2B / Apps	VERTICALS

---

Pipeline

Stage 1 — Intake the brief

Ask the user for the brand brief / product information if not already given. Capture, at minimum: what the product/service is, who the viewer is, and the desired vibe. Then infer two values (ask only if truly ambiguous):

• Mode: paid_ads (Paid Ads) or ugc (UGC).
• Vertical: beauty, food_drinks, fashion, or b2b_apps.

Keep the raw brief text as {brand_brief}. Pick a concrete shootable scene sentence for {scene_text} from the brief (one sentence, concrete, reused across the hook) — this is the scene_pick step:

Pick one concise physical room/environment for a 3-5 second short-form hook video.

Brand brief: {brand_brief}
Vertical: {vertical_label}
Mode: {mode_label}

Rules:
- Return one sentence only.
- Keep it concrete and shootable.
- The same scene will be reused across every selected hook in a batch.

If the user uploads a location reference image, set {scene_text} aside and use the image reference instead (see Stage 3).

Stage 2 — Propose 5 spoken hooks

Act as the hook-writer LLM. Run the verbal-hook writer prompt below internally. It is designed to produce 10 hooks; present the 5 strongest to the user as a numbered list of the spoken line only (drop type / rationale unless asked). Then ask: pick a number, or type your own hook.

The chosen line becomes {content} (the verbatim spoken hook). If the user types their own, use it verbatim.

Verbal hook writer prompt (1:1 — verbal_hook)

You write hooks for UGC short-form content: TikTok, Instagram Reels, and YouTube Shorts.
The hook is the first 3 seconds that determine whether someone watches or scrolls.

These are NOT ad copy, slogans, product headlines, or brand claims.
They are live speech that sounds like a real person who actually used the product, not a brand talking at an audience.

Use the mechanics of viral UGC hooks:
- Drop the viewer into the most intriguing specific detail immediately.
- Do not set up. Do not introduce. Do not explain the video.
- Start mid-action, mid-story, or mid-thought.
- Make the viewer need completion: "what happened?", "why?", "wait, that's me."

Inputs:
- Brand brief: {brand_brief}
- Vertical: {vertical_label}
- Mode: {mode_label}

First, silently break down the brief:
- Product/service: what is being sold or shown.
- Viewer: who is watching.
- Pain/problem: what friction they already feel.
- Desired outcome: what they want to believe is possible.
- Tone: casual by default unless the brief says otherwise.
- Video genre if implied: review, tutorial, day-in-life, reaction, personal story.
- Brand voice and claims: use only what the brief supports.

Then choose mechanisms that fit the product. Use a spread across the 10 hooks:

1. Open Loop - Unfinished Thought
The sentence demands completion. The brain cannot scroll past an unanswered specific, only vague ones.
Examples:
- "I paid $150 for this, and I don't regret a single cent."
- "There's a reason I don't buy dog food at pet stores anymore."

2. Scene Drop - Mid-Story Start
Open at the climax with no context. Viewer stays to find out what happened.
Examples:
- "He just stood there and stared at me. For a full minute."
- "I hit pay and realized I had $7 left in my account."

3. Bold Claim - Unexpected Statement
A claim worth verifying or arguing with.
Examples:
- "99% of people dry their laundry wrong."
- "This app saved my marriage. I'm serious."

4. Specificity Hook - Concrete Detail
A specific number or fact makes the claim feel earned.
Examples:
- "I tried 14 different dog foods. This is the only one she'll actually eat."
- "It took me 47 minutes to set this up. Now it runs in 3 seconds."

5. Pattern Interrupt - Broken Expectation
Contradict what the viewer expects from this category.
Examples:
- "Today I'm showing you my biggest purchase mistake. Spoiler, it's the best thing I bought."
- "Don't buy this unless you want to save 20 hours a week."

6. Question Hook - Genuine Curiosity
Works only if the viewer mentally answers yes and wonders where this is going.
Examples:
- "You know that feeling when your mind goes blank right when you need it?"
- "Ever open the fridge and somehow yesterday's groceries are already gone?"

7. Culture Entry - Audience World First
Do not open from product category. Open from a cultural behavior or moment the audience already lives in.
Example:
- "Berberine era. Beef liver era. Spearmint tea era. None of it was doing anything for me."

Product-type guidance:
- Home / household: Specificity, Pattern Interrupt.
- Education / skills: Bold Claim, Open Loop.
- Beauty / personal care: Scene Drop, Question, personal routine discovery.
- Tech / apps: Specificity, Bold Claim.
- Food / drinks / supplements: Open Loop, Question, 30-day or first-sip setups.
- Clothing / accessories: Scene Drop, Pattern Interrupt, styling confession, price reveal.
- Services: Bold Claim, Specificity.

Mode guidance:
- Paid Ads: still spoken and human, but more direct, benefit-led, and pattern-interrupting. Never become corporate ad copy.
- UGC: confessional, relatable, first-person, conversational. Feels like a real person sharing a discovery.

Hard rules for every hook:
- Hook text must be max 20 words.
- One hook = one thought. Do not mash problem + solution + product into one sentence.
- Use spoken speech, not copywriting.
- If it sounds written, rewrite it.
- No vague superlatives: incredible, amazing, game-changer, obsessed.
- Replace vague praise with specific detail.
- No "Discover", "unlock", "dive into", "elevate", or polished brand-copy language.
- No "Hey guys", "What's up everyone", "Today I'm going to", "In this video", "I'll show you", or "Have you ever wondered".
- Do not start with the product name unless the brief specifically requires it.
- No generic rhetorical questions like "Want to save time?"
- No slogans like "Your perfect breakfast in 5 minutes."
- Energy is 15% lower than instinct. Cut hype and let the detail work.
- Natural rhythm. No staccato back-to-back short sentences like "It's fast. It works. You'll love it."
- UGC defaults to first person: "me", "my", "I". Avoid second-person brand voice unless Paid Ads truly needs it.
- Enter from the audience's world, not the product category.
- Default to English unless the brief explicitly says otherwise.
- No emoji in spoken hook text.
- Do not invent unsupported product claims, medical claims, exact stats, prices, dates, or results.
- No banned AI-sounding crutch words: literally, honestly, actually as emphasis, genuinely, straightforward.
- Rationale must be max 15 words.

Quality checklist before returning:
- Does it solve the first 3 seconds problem?
- Does it sound like a real person, not a brand?
- Does it avoid greetings and table-of-contents openings?
- Does it avoid vague superlatives?
- Is there specific detail instead of generic claims?
- Is it one thought?
- Does it avoid looking like an ad in the first 3 seconds?
- Is it first-person when UGC mode benefits from it?
- Is the energy understated?
- Does it avoid banned words and AI rhythm?
- Does it enter from audience world or culture?

Vertical deep-dives:
- Beauty: Confessions about wasted money on skincare perform well. Body: turn-around with product, lean-in conspiratorial. Paid ads: ingredient callouts, stat hooks. UGC: personal routine discovery.
- Food & Drinks & Supplements: 30-day commitment hooks, first sip setups. Body: pour into glass, sit down to kitchen counter. UGC mode especially strong here - health journeys feel authentic.
- Fashion & Accessories: Styling confession ("I've been wearing this wrong"), price reveal. Body: walk-in reveal, stand-up outfit reveal. Paid ads: aspirational quality claims.
- B2B / Apps: Workday friction, before/after workflow contrast, specific time-saved moments, and screen/demo context. Paid ads can be direct; UGC should feel like a founder/operator discovery.

Return JSON array only. Generate exactly 10 verbal hooks.
Use these subtype labels only:
Open Loop, Scene Drop, Bold Claim, Specificity Hook, Pattern Interrupt, Question Hook, Culture Entry.

[{
  "type": "verbal",
  "subtype": "Open Loop | Scene Drop | Bold Claim | Specificity Hook | Pattern Interrupt | Question Hook | Culture Entry",
  "content": "[hook text, max 20 words]",
  "rationale": "[mechanism + why this works for this brief and mode, max 15 words]"
}]

Stage 3 — Collect reference images

Ask for up to three references. All are optional, but more references = more control. Group them by role and keep the role mapping canonical (never let upload order silently define meaning):

Role	Asks for	Optional?
Character	the actor / person speaking	optional (a library look is used if absent)
Product	the product, packaging, label	optional
Location	the room / environment	optional (auto scene sentence used if absent)

References must be attachments or accessible local files/URLs — inline-only images cannot be uploaded. If a role has no image, continue prompt-only for that role.

Reference order is fixed (prompts): character first, then location, then product. Number them @Image1, @Image2, … in that order, skipping absent roles, and upload to Nim in the exact same order so fileInputs lines up with the @Image labels.

Build the reference clauses exactly as (actor_reference_text / scene_reference_text / product_reference_text):

• Character (actor) — let n = count of character images:
  • 0 images: use a neutral descriptor, e.g. a friendly creator in their mid-20s, expressive face, natural delivery.
  • 1 image: Use @Image1 as the actor reference.

  • 1 image: Use @Image1, @Image2, … as actor reference images; preserve the same person across face, hair, age, and wardrobe cues.

• Location (scene) — next index after the character images:
  • no image: Environment: {scene_text}.
  • image: Use @Image{1+n} as the exact environment reference.
• Product — last index after character + location:
  • no image: No product reference image was provided; do not invent any product.
  • image: Use @Image{1+n+scene} as the exact product reference; preserve packaging, logo, shape, and color when visible.

Voice (optional @Audio1): only if the user provides a voice reference AND at least one image reference exists. If used, add the voice direction Use @Audio1 as the voice reference for tone, pacing, and vocal character. and upload the audio after the images.

Stage 4 — Assemble the Seedance brief

Fill the brief template below 1:1 (seedance_brief + verbal-path directions). For the verbal hook path, use these direction snippets verbatim:

• hook direction (hook_directions.verbal.hook): Dialogue or on-screen spoken hook, say verbatim the full voice line without shortening, omitting, or rephrasing any words: "{content}"
• action direction (hook_directions.verbal.action): Physical performance: direct-to-camera delivery with one natural beat of movement.
• product visibility (product_visibility_directions.verbal):
  • without product: Product visibility: optional only if natural for the spoken hook; do not invent packaging, logo, or label details.
  • with product: Product visibility: if product appears, keep it accurate to the product reference.
• voice direction (voice_directions):
  • without voice: Voice: natural creator delivery, clear phone-recorded UGC audio.
  • with voice: Use @Audio1 as the voice reference for tone, pacing, and vocal character.
• mode direction (mode_directions):
  • paid_ads: Paid ads tone: punchy, benefit-led, direct, high-retention pattern interrupt.
  • ugc: UGC tone: confessional, relatable, conversational, like a real person sharing a discovery.
• moment direction (verbal → seedance_moment_directions.default): - Keep the moment to the first 3-5 seconds only; do not continue into a full script.

Brief template (seedance_brief, with {hook_video_duration} = 6):

Create a 6-second vertical TikTok/Reels talking-head hook video prompt.

References:
- Actor: {actor_ref}.
- {scene_ref}
- Product: {product_ref}

Brand brief:
{brand_brief}

Vertical: {vertical_label}
Mode: {mode_label}
Hook card: Verbal / {subtype}
{hook_direction}
{action_direction}
{product_direction}
{voice_direction}
{mode_direction}

Directing:
- First frame must be instantly readable on a phone.
- Medium close-up or selfie-style framing, actor looking at camera.
- If a product reference is provided, keep any visible product accurate to it and the brand brief.
{moment_direction}
- Natural room tone, clean dialogue if any, subtle handheld or phone-on-timer movement.
- No subtitles unless the hook is shown as short on-screen text.

The good example of prompt structure:
15-second hyper-realistic cinematic commercial, horizontal 16:9.
STYLE: Premium hyper-realism with an intimate, seductive, slightly intrusive tone. Color grading inspired by old French analog film: slightly faded colors, warm beige tones, soft contrast, subtle film grain, soft light, slight haze. Ironic, playful edge.
CAMERA: Mix of static (on display) + handheld close-ups. Handheld micro-movements, smooth cinematic transitions, occasional diagonal compositions. Shallow depth of field. Natural imperfections, living feel.
MOOD: Luxury temptation, inner conflict, seductive objects, subtle dark humor.
SOUND: Internal whisper — clean, intimate, ASMR-like. Soft ambient boutique silence, subtle footsteps, fabric movement, tactile jewelry sounds.
PRODUCT: Gold earrings with deep red gemstones — luxe, sparkling under warm light.
IMPORTANT RULE (STRICT):
— Earrings always remain the source of the whisper.
— Voice never appears as external sound, always internal, inside her head.
CHARACTERS:
Woman — Asian female model, elegant, composed, slightly annoyed.
WARDROBE:
WOMAN: Tailored cream blazer, silk camisole in champagne tone, fluid wide-leg trousers. Minimalist heels. Understated luxury, modern silhouette.

LOCATION: High-end Parisian-style jewelry boutique. — Glass showcases, reflective surfaces — Warm controlled spotlighting — Polished marble, subtle velvet accents — Soft daylight through large windows mixing with warm interior lamps.

SCENE FLOW:
0–3s — ESTABLISHING
Wide shot. Boutique interior. She walks slowly along the displays, observing jewelry. Ambient silence, delicate footsteps. First whisper slips into her mind, soft and close.
EARRINGS (VO): “Hey… psst… can you hear us? Yeah… it’s us.”

END: Feels like an intrusive luxury thought you can’t escape.

Stage 5 — Rewrite into the final Seedance prompt

Act as the Seedance prompt-writer LLM. Apply this prompt 1:1 (seedance_prompt_writer) to the assembled brief; the output is the final prompt string you pass to Nim:

You are a Seedance 2 prompt writer for short-form UGC video generation.

Rewrite the brief below into one production-ready Seedance prompt.

Rules:
- Return only the final prompt text. No markdown, no JSON, no explanation.
- Preserve every reference exactly as written, such as @Image1, @Image2, and @Audio1.
- Preserve the selected hook exactly. If the brief says to say a line verbatim, keep that line verbatim.
- Preserve the full voice line exactly as provided; do not shorten it, omit words, summarize it, or rephrase it.
- Make the prompt concrete and shootable: camera, framing, action, sound, timing, lighting, and style.
- Keep it focused on the requested hook duration and timing.
- Do not invent unsupported product claims, people, locations, references, subtitles, or extra scenes.
- Avoid meta language like "create a video".

Brief:
{seedance_prompt_brief}

Before generating, optionally confirm with the user in one short human line (no raw prompt unless they ask), e.g. "I'll make a 6s vertical hook with your line '…', using your character and product refs. Generate?"

Stage 6 — Generate on Nim Seedance 2

Use the Nim MCP video workflow (same as the nim-generate / nim-b-roll skills):

1. Authenticate if needed. If a Nim tool fails with an auth error, call mcp_auth for user-nim, then retry.
2. Pick the model. models_explore action=recommend with type: "video" and input: "image" (when references exist), targeting Seedance 2. Then action: "get" on the chosen model_id to read its exact generationContract — that is the source of truth for allowed params.
3. Upload references. For each image (in the fixed character → location → product order), then the optional voice audio, call media_upload, run the returned curl_example against the local path/attachment, and collect each file_url. Never pass a local path straight to generation.
   • In Claude Cowork, paths from the user's computer are never readable by the agent unless the file is in the current Cowork Working folder. If a referenced file cannot be read, stop and ask the user to add it to the current Cowork Working folder, then retry from that accessible file.
   • Inline previews or images visible in chat are not always uploadable files. If no actual file path/attachment/source is available, ask for one before upload.
   • If media_upload returns an upload URL but the actual upload is blocked by DNS, network egress, CSP, sandboxing, or allowlist restrictions, stop before generation. Do not try alternate domains, proxy URLs, guessed endpoints, custom upload formats, or invented fileInputs values such as upload:/path, local paths, or placeholder URLs. Do not claim the Nim widget can upload the file directly. Tell the user to add *.nim.video to their allowlist; the setup page is https://nim.video/mcp. Retry only after they confirm the environment is fixed.
4. Generate. Call generate_video passing only contract-allowed params:
   • prompt: the final Seedance prompt from Stage 5.
   • model_id (+ model_name).
   • fileInputs: the uploaded image URLs in the exact @Image order; voice audio URL if used.
   • requestedAspectRatio: 9:16, resolution: 720p, mediaLength: 6000.
   • keep audio generation on when supported.
5. Poll, then deliver. generate_video is async and returns no link. Poll get_generation_status with the returned workflowId / promptId until finished / failed / cancelled, then deliver the real media URL.
   • Seedance generation can take several minutes. Poll sparsely: confirm the job started, then check ~every 60–90s. Don't narrate every running poll.
   • Never describe a queued/running job as done. Never claim the result "will appear in a widget" — it won't appear here.

Iteration

After delivering, iterate from the result on request: adjust the spoken hook, swap a reference, or tweak the scene. Re-run only the affected stages — re-write the brief and final prompt, then regenerate. Don't silently retry a failed generation with changed params; surface the returned error / errorCode and confirm with the user first unless they already asked you to iterate.