| name | direct-street-interview-video |
| description | Design natural vertical street-interview and walk-and-talk video prompts for MiniMax H3 or Seedance 2.0. Use when a user wants spontaneous documentary energy, a referenced recurring person, first-person handheld follow, concise dialogue, moving daylight, believable street parallax, or a consent-led interaction that must fit 4–15 seconds without unreachable beats. |
Direct Street Interview Video
Create an original street interaction from a reusable documentary mechanism. This is a non-official, user-contributed Skill; it is not a MiniMax or ByteDance preset.
Required reading
- Read references/summary.md for style, fit, usage, and source limitations.
- Read references/template.md for invariants, variable slots, ablations, transfer tests, and repairs.
- Read the target model template before compiling:
- Consult references/source-analysis.md only when explaining why a repair exists. Do not reproduce its people, wardrobe, location, or dialogue.
Intake
Collect or infer only what is needed:
- target model, task type, duration from 4 through 15 seconds, and aspect ratio;
- the interaction goal, location, subject role, interviewer role, relationship, and ending;
- every generation-time image or video and its narrow role;
- exact approved dialogue, speaker order, and language;
- whether physical contact is explicitly wanted and appropriate;
- the one primary camera behavior, ambient sound, and any requested music or visible text.
Treat the analyzed demonstration video as research evidence, not as an automatic generation reference. Never fabricate a reference asset or dialogue.
Build the scene
- Start with one readable social premise: approach, recognition, question, demonstration, compliment, request, or compact reveal.
- Give the on-camera subject one stable visual identity. With reference media, bind two or three distinctive anchors and preserve wardrobe continuity.
- Use one continuous route unless a new shot is causally necessary. The default camera is first-person handheld backpedaling or side-walking at conversational distance.
- Make the relationship change visible through distance, gaze, pace, hand gesture, body angle, or a shared object. Do not substitute random camera motion for performance.
- Let the street move around the subject: foreground occlusion, background parallax, passing light and shadow, and small exposure adaptation.
- Keep dialogue subordinate to visible action. Allow each line to finish before the next speaker begins.
- End on a visible resolved state: permission granted, pose found, object shown, route chosen, reaction landed, or conversation naturally released.
Feasibility gate
- Every timestamp or later beat must occur before the actual target duration.
- For about 10 seconds, prefer two or three short spoken turns and one continuous event chain.
- For about 15 seconds, prefer no more than four or five short turns unless speech is intentionally rapid and the user supplied it.
- Remove a shot before compressing dialogue into unnatural speed.
- Do not place a new setup exactly at the video endpoint.
- Do not claim 4K, HDR, film stock, lens metadata, or platform provenance unless the user requested or supplied it.
- Default to verbal or gestural positioning. Add touch only when the user explicitly wants it and the staging is consensual, motivated, and physically clear.
Compile for MiniMax H3
- Select T2VA, I2VA, FL2VA, L2VA, or Ref2VA from the actual generation-time inputs.
- Use the three-field Base structure when no reference relationship is needed.
- Use the six-field Ref2VA structure when a supplied person, wardrobe, voice, scene, or movement reference must be tracked.
- Write official-strict rewrite prose in English. Preserve approved dialogue only in
<d>[Language]...</d> and identify speakers with (S1), (S2).
- Start with
[Shot 1] and no timestamp. Later shots use consecutive labels and increasing timestamps below the target duration.
- Keep dialogue only in the detailed visual timeline; do not repeat it in the soundscape.
Compile for Seedance 2.0
- Write in Chinese by default and choose a task from the actual inputs.
- Use one compact paragraph for a continuous walk-and-talk; use consecutive
镜头N only when there is a real cut or event boundary.
- Bind references as
角色A@图片1 or equivalent and give every asset one explicit role.
- Put dialogue in
{...}, physical sound in <...>, music in (...), and requested visible text in 【...】.
- Do not use H3 fields, H3 labels, speaker IDs, or exact per-shot timestamps.
Validation
Before returning the result, verify:
- duration, speech, and action are jointly feasible;
- every asset reference resolves and every declared asset has a purpose;
- the subject, wardrobe, environment, route, light direction, and hand/object state remain continuous;
- camera motion follows the interaction and does not conflict with itself;
- dialogue order, lip movement, ambience, and optional music do not duplicate or contradict each other;
- the execution changes subject, place, wardrobe, dialogue, route, framing, and ending instead of copying the demonstration surface;
- no paid generation, upload, ComfyUI import, or node modification is performed without a separate request.
Deliver
Return:
- one-sentence concept;
- task and asset-role map;
- a compact feasibility note;
- the MiniMax H3 prompt;
- the Seedance 2.0 prompt;
- a short continuity and anti-copy audit.