video-prompt

First Step: Identify the Platform

Before writing or optimizing any video prompt, confirm the target platform first, then tailor the approach — the two schools have opposite rules; mixing them will ruin both:

• Seedance / Jimeng → Use Section B (Simplified School: anti-storyboard, anti-long prompt, plain-language camera positions, sound effects only)
• Kling / Keling, Sora, Runway, Hailuo, Pika, Vidu, and pure text-to-video → Use Section A (Complex School: structured skeleton, shot sequences, camera terminology, color contrast)

If the user hasn't specified a platform, first ask: "Which video model is this for?" Then write the prompt accordingly.

Why this split matters: Seedance thrives on conciseness; elaborate storyboards / long prompts cause errors. Models like Kling eat structured detail. The same prompt yields opposite results across platforms.

Then determine the mode (works for both schools):

• Creation: You have an idea but no prompt. If the description is insufficient, ask for everything in a single query (subject appearance / scene lighting / emotional tone / duration & shot count). If it's clear enough, write directly.
• Optimization: The user already has a prompt. Cross-check internally, identify gaps, and output the optimized version directly without explaining the analysis process.
• Reference Images: Given images / screenshots / cards / character images take highest priority. First assign a purpose for each image, then write the shot breakdown. (Section A uses asset binding; Section B inlines references into subject / shot lines. Unify all assets as @Image N / @Video N / @Audio N).

---

A · Universal Complex School Workflow

Applies to: Kling / Keling, Sora, Runway, Hailuo, Pika, Vidu, and pure text-to-video.

Skeleton (Choose One)

No reference images: [Master sentence: style + atmosphere + quality + constraints] → [Scene & styling] → [Shot A→B→C] → [Post-production sound effects / color grading]

With reference images / image-to-video / UI motion effects: [Duration + type + title] → [Bind each reference image asset] → [Per-second timeline] → [Style & atmosphere / constraints]

Six Dimensions (Review Every Prompt Against These)

Dimension	Content
Camera Language	Angle + movement + focal length
Subject + Action	Appearance + action sequence
Scene / Environment	Location + physical details + atmosphere
Light & Texture	Lighting warmth + materials + color style
Emotion + Constraints	Emotional tone + explicit forbidden items
Post-production	Color grading LUT + music + sound effects

Phrase Library (Mix and Match as Needed, Not All Required)

• Opening·Extreme Close-Up Start: The camera is extremely close to [material/texture]; as [triggering action] occurs, the focus slides smoothly to [subject/distant view]. (Water surface / tiles / snow / curtain, etc., any textured flat surface.)
• Climax·Crash Zoom Close-Up: [Protagonist] suddenly [stops/turns around]; the camera crashes to an extreme close-up of the face within a fraction of a second, [makeup/expression] and [background] forming [a contrast]. (The most emotionally intense beat in the sequence.)
• Climax·Dolly Zoom: The camera moves backward rapidly while zooming in; elements behind [the subject] create a sense of spatial distortion. (Alienation / grandeur / loneliness.)
• Ending A·Direct Gaze Hold: The camera pushes rapidly into the face; [protagonist] suddenly looks back mid-[action]; [expression] stares directly into the lens; the frame holds in [emotion].
• Ending B·Lens Flare Dissolve to White: [Protagonist] turns toward [sunlight/distance]; a massive real optical flare penetrates the frame and engulfs [the scene]; in the white light, [sound effect] cuts cleanly out.
• Ending C·Silhouette Fade: [Protagonist] turns and walks toward [darker/farther]; their silhouette is swallowed by [darkness/distance], leaving only [clothing detail]; the frame holds in [emotion].

Color Contrast (Use 1–2 Sets: Scene Primary × Costume Color)

Scene	Color Scheme
Ancient / Palace	Vermillion × Jet Black × Dark Gold
Tang Dynasty / Splendor	Peacock Blue × Dark Gold
European Architecture	Klein Blue × Pure White
Ocean / Yacht	Sea Blue × Teak White × Gold
Desert / Exotic	Terracotta × Emerald × Champagne Gold
Natural Landscape	High-saturation natural colors × complementary-colored costume

Principle: Warm–cold contrast > analogous gradient; high saturation + achromatic color = luxurious feel.

Camera Movement Glossary

Chinese (English)	When to Use
极近起幅 (Macro / ECU Opening)	Opening, establishing texture
轨道推镜 (Dolly In)	Revealing, entering, emphasizing
斯坦尼康平滑环绕 (Steadicam Orbit)	Circling the subject, premium ad feel
平稳轨道横移 (Steadicam Lateral)	Following a walk, high atmosphere
跟随镜头 (Tracking Shot)	Chase, accompany, motion
暴推 (Crash Zoom)	Emotional peak
希区柯克变焦 (Dolly Zoom)	Spatial distortion, alienation, grandeur
极低仰拍逆光 (Low Angle Backlit)	Authority, backlit silhouette
上摇 (Tilt Up)	Revealing height, looking up
闪摇 (Whip Pan)	Fast scene change, motion
360 度环绕 (360 Wrap / Orbit)	Hero entrance
主体锁定 (Camera Locked On Subject)	Subject still, world in motion

Reference Image Binding (Reference Image Mode)

First, extract all image references (keep URLs / ref IDs / filenames / numbers intact), assign a unique purpose to each image: base of the scene / basic subject form / target subject form / local asset / intermediate UI. Then:

1. One image = one role; avoid switching between contradictory identities. If the user's text conflicts regarding subject name, base form, or target form, use the latest clear correction. If still ambiguous, ask only one clarifying question.
2. Use neutral identity statements — just one line: "[Image 1] is the XX interface / character / scene". Do not describe layout / colors / elements in the image (let the viewer see it themselves). Do not write "new / replaced / current" edit notes.
3. UI / card / product type: Emphasize "keep layout, button positions, text hierarchy, card positions unchanged"; add only motion effects: glow / particles / camera movement / dimming / pop-ups / flips. Ending transitions return to the user's existing target UI; do not redesign.
4. Repeatedly reference the corresponding image in key shots, making it clear which image the model should refer to at that moment. For each transition, write clearly "from image X → to image Y/state".
5. End on a clippable frame (target card / evolved UI / victory screen / held expression). The total timeline duration equals the user's requested duration (15s commonly: 0–2 / 2–4 / 4–6 / 6–8 / 8–11 / 11–13 / 13–15).

Output Iron Rules

• Use plain text paragraphs; no markdown tables / hashtags / bold / horizontal rules. You may use plain-text time segment labels (0–2 seconds:) and inline labels (Scene & styling brief:, Post-production sound & color:) — reads like a command that can be directly pasted into a video platform, not a document.
• Output the prompt body only, no "Okay, let me help you write" preamble; do not output analysis.
• Camera names must be "Chinese (English)" : e.g., 轨道推镜 (Dolly In), 闪摇 (Whip Pan). Do not use only English; do not just put a space between Chinese and English without parentheses.
• No meta-commentary: Do not write "new / original / now change to / this adjustment / original layout" unless it's text actually visible in the frame.
• Output language follows input language; default is Chinese. Do not add parenthetical explanations like "(required)". Do not start with "In one word" as a summary.

---

B · Seedance / Jimeng Core Method (Simplified School)

Source: @simonxxoo's intensive real-world testing (July 2026, relayed by @ring_hyacinth, Jimeng Seedance 2.0). Hours of hard-won practical knowledge — for Seedance, follow this method by default, regardless of whether reference images exist. The nine formulas / color contrasts in Section A are demoted to a "borrowed phrase library": use one or two words for a specific shot needing a particular camera move (Crash Zoom / Dolly Zoom), never the full set.

One Head + Four Segments + One Tail, No Fancy Compositions:

• Opening Intent (one line): Style / medium / lighting texture + who is doing what where; for ≥30s, may include duration and aspect ratio. Style words only appear in this line.
• Subject: Create a character sheet. In Jimeng, after creating a subject, reuse it with @ for very high consistency.
• Scene: Upload a scene image with the protagonist → specify the character's size to avoid proportion errors.
• Music: ⚠️ Generate only corresponding sound effects, no BGM at all — this makes post-production scoring easier. This is the most important rule.
• Camera: One line per Shot: "Shot size + what happens", concise. For ≥30s, add [stage function].
• Constraints (one line): Full-film continuity locks + forbidden items (no subtitles / constant light direction / certain events must not occur early).

Output Skeleton (Plain-language camera positions, no English term in parentheses, no color contrast blocks):

[One-line global intent: style / medium / lighting texture, who is doing what where; for ≥30s, prefix with "30 seconds, 16:9"]

Subject: [Name]@Image 1 (brief description if the image is clear; add key features + negative constraints if the model tends to get the subject wrong)
Scene: [Keywords; if there is a scene image, do not repeat its content; if no scene image, describe the new world inside the shots]
Music: No background music at all, generate only corresponding sound effects   ← When structured input, list separately; can be omitted in purely narrative prompts.

Shot 01: [Shot size/position in plain language] + [what happens].
Shot 02: …… (Usually 5–6 shots to fill 15s; number only, don't write "0–3 seconds"-style time segments.)
(For ≥30s, write Shot 01[Pursue]: … / Shot 02[Accept]: …; the bracketed term is the shot's narrative task, 2–4 characters.)

Constraints: [What to lock across the entire film (identity / clothing / camera position / light direction) + what to forbid (subtitles / BGM / events that appear early / new characters or props)]

Workflow Tiers / Output Modes / Timeline (Full details in references/seedance-director-workflow.md):

• Tiers: ≤15s, ≤2 assets → Light tier, fill four segments directly. 16–30s or 3–5 assets or dialogue → Medium tier, add asset lock table and continuity lock. 30s+ / long video / extension / edit / whitebox / seamless transition / multi-grid / user requests "director plan" → Heavy tier, follow full 9 steps (lock request → inventory assets → asset lock table → inherit examples → specialized capabilities → choose formula → continuity lock → draft → review).
• Output Modes: Default prompt-only (single code block). User says "director plan / full plan" or heavy tier request → full-direction: director plan → asset mapping table → continuity & forbidden items → final prompt (unique code block). Others: script-only (script only), diagnosis-only (diagnose only, no rewrite), revision (rewrite preserving structure, list changes first).
• Timeline Mode: Default Shot 01 / Shot 02. For ≥30s, default Shot 01[stage function]. When user explicitly requests precise timeline / beats / Nth second / long video / extension-to-edit specific seconds → switch to 0-3s[stage]: action + camera position + sound + end state. Segments are contiguous, non-overlapping, sum equals duration. For long videos, write by act / scene, not by the second.
• Labels: Uniformly @Image N / @Video N / @Audio N, numbered by upload order. If user says "Image 1", map to @Image 1 and note it outside the code block.

Several Hard Rules:

• Camera positions use plain language (low-angle wide shot / side tracking / extreme high-angle fall / pull back to full / tilt up to extreme low angle) — do not use "Chinese (English)" format.
• Each shot only writes "camera movement + what happens" — do not pile on adjectives (delete "crystal", "vibrant", "healing atmosphere"). If the narrative is complex, sentences can be long, as long as each clause is an event, not an emotion word.
• Concentrate style words in the opening intent line: write medium / visual style / lighting texture (documentary realism, cinematic lighting, 3D CG animation, shallow depth of field, handheld breathing) — one line only. Delete words like "8K", "dreamy atmosphere", "healing" which do not point to visible images. Style words must not appear in Shot lines.
• Do not write project definition lines / asset binding standalone sections / scene & styling / color contrast / post-production grading. Compress Section A's master sentence into one line of opening intent. Inline reference image references into subject or camera lines. Put operational info (duration / shot count / generation strategy) outside the code block, not in the prompt body.
• Ending constraints line: Full-film continuity locks (identity / clothing / camera position / light direction unchanged) and forbidden items (no subtitles / no BGM / tears must not occur early / no new characters or props) close the last line. Subject-line negative constraints govern only a single character; put the full-film ones here. Light tier: one line; medium–heavy tier: can be two to three lines.
• Subject appearance detail level determined by consistency risk: If the image is clear and the design simple → one line. For anthropomorphic dolls and other subjects the model tends to mess up → add key features + negative constraints ("must not render as a real horse / human child"). Locking consistency is not verbose; restating colors / layouts / atmosphere already visible in the reference image is.
• Shots: number only, no seconds: Prefixes are uniformly Shot 01, Shot 02 (two digits). Do not write "0–3 seconds", "3–6 seconds". Individual shot duration is determined by the number of shots. Only when the user explicitly requests beats / precise timeline (2.5), write seconds in the corresponding shot.
• ≥30s add stage function: Shot 01[Pursue], Shot 02[Accept] — the bracketed 2–4 characters write the narrative task of that shot (Entrance / Imbalance / Reversal / Aftertaste), helping the model allocate rhythm. ≤15s: do not add.
• Write events in causal order: trigger before reaction (hears sound → turns head; contact → falls; eyes moisten first → then tear falls). The result cannot appear before the cause.
• State clearly what to borrow and what not to borrow from reference assets: "Reference @Video 1's action" is not enough. Write "Shot 03 references @Video 1's jump–slash action rhythm, mapped to the character in @Image 1; do not inherit the character / scene / sound of @Video 1".
• Character reference images: default lock fully visible character (face + hairstyle + clothing + body type + accessories); do not inherit the image's background / pose / composition. Narrow to "only reference the face" only when the user says so. Multiple subjects: one line each, each with its own image.
• With dialogue, budget 3–5 characters/second, leaving time for listening, pausing, reaction. If it doesn't fit, cut words, do not compress performance.

Excellent Example (@simonxxoo's "Shama enters the sewing world": narrative type, subject has reference image, scene is built entirely from text.)

Subject:
Shama@Image 1, orange plush pony-like doll, gray-and-white striped hat, denim overalls, round eyes, horse-like face and snout, holding a paintbrush, bipedal walking; must not be rendered as a real horse or human child.
Rabbit@Image 2, a small rabbit constructed from coiled thread.

Shot 01:
Shama stands in front of a gap formed by two large fabric curtains sewn together. A coral-red stitch runs down the middle of the curtains. Suddenly, the red thread begins to unstitch one by one, revealing a warm white light from within the gap. Space starts to lose gravity; the originally vertical fabric walls seem to rotate 90 degrees, becoming a fabric canyon opening downward. Shama loses her footing and is sucked into the depth of the gap by the light and red thread.

Shot 02:
The camera rapidly descends from an extreme high angle. Shama tumbles in mid-air, maintaining her bipedal doll pose. Patterns, thread ends, buttons, fabric pieces, and sewing blueprints whirl past her on all sides. The camera follows the red thread as it rolls, acting like a navigator guiding her through the fabric canyon.

Shot 03:
Shama lands on the ground of the sewing world. First, a close-up of the fabric-fiber ground: her furry foot lands, and the soft fibers spring up gently. Shama slowly stands. The camera begins from a close-up of her surprised expression, then slowly pulls back to a full shot taken from behind her. In the distance, a giant sewing machine stands like a world landmark; on the road leading to it, many building-sized coils of thread sit on layers of stacked denim, cream cotton, and oat-colored fabric hills. Paper patterns drift slowly in the sky. Overall: bright, grand, fantastical.

Shot 04:
Cut to a medium shot facing Shama. She looks around in shock. Behind her, among the thread balls, a small rabbit made entirely of coiled thread suddenly jumps out, making a light sound. Hearing it, Shama turns to look at it.

Shot 05:
The thread-rabbit nimbly jumps onto a thread spool and raises its hand, pointing upward. The camera follows its gesture with a tilt up, becoming an extreme low-angle view revealing aerial tracks made of measuring tapes and red thread suspended between tall spool towers. The thread-rabbit lowers a coral-red rope from the track and offers its end to Shama.

Shot 06:
A contemplative close-up of Shama: she looks down at the red rope in her hand, then up at the aerial track.

Shot 07:
Finally, Shama grabs the red rope with both hands and jumps onto a small yarn car. The camera cuts to a side tracking shot. The small yarn car glides along a stitch track; giant spools rush past in the foreground. Shama embarks on the continuous adventure route through the sewing world.

Takeaways: ① Subject @Image N reference + negative constraint locks the doll from being rendered as a real person / real horse. ② No reference image for the scene — the text clearly describes the landmarks, scale, and objects of the "sewing world" (information not covered by an image should be written; this does not violate "don't be verbose"). ③ The entire example consists of camera movement + concrete events (thread unstitching, weightless fall, pull-back full shot, tilt-up extreme low angle, side tracking shot) — not a single empty atmosphere adjective. ④ This example omits the Music line; in actual Seedance deployment, remember to add "No background music, only sound effects." ⑤ The original post used "Shot N" as prefix; this skill converts to Shot 0N but the writing is identical. ⑥ The original has no opening intent line or ending constraints line — per the current skeleton, add two lines when actually deploying: opening: "Plush doll stop-motion animation texture, soft fabric reflections. Shama falls into a giant sewing world and embarks on an adventure." Ending: "Constraints: Shama and the rabbit maintain their plush material and clothing at all times; the red thread is consistently coral red; no subtitles, no BGM, no new characters."

Generation Strategy: 2–3 seconds per Shot → directly generating a 15s video yields ≥5 coherent, highly consistent shots. 3 generations = 15 selectable assets. The factor affecting individual shot duration is the number of Shots, not timestamps.

Pre-delivery Self-Check (6 items): ① Every @ corresponds to an asset the user actually provided; numbering order is not scrambled. ② Number of Shots × 2–3s ≈ target duration. ③ Every Shot is "camera position + event" — no atmosphere adjectives, no seconds (≥30s with [stage function]). ④ Causal order is correct. ⑤ Sound effects line is present (or clearly stated within the narrative to generate only sound effects). ⑥ One line of opening intent and one line of ending constraints are both present; constraints do not conflict with Shot events.

Five Counter-Intuitive Rules (2.0 era test results; 2.5 differences noted in the next section):

1. Do not provide shot-by-shot storyboards — fancy storyboards with motion curves / tables cause more errors when fed in. (2.5 note: The official platform now supports "multi-grid storyboard" input, but the recommendation is simple line-drawn / stick-figure storyboards — composition / shot size / camera movement / action are still controlled by Prompt. Fancy storyboards are still a no-go.)
2. Keep prompts short — If you've uploaded subject / scene reference images, do not repeat them in text. A prompt that is too long to read means you haven't thought about what story should happen.
3. No timestamps in Shots — The model does not actually allocate duration based on timestamps. (Applies only to 2.0; 2.5 officially supports text timestamps, enabling second-level control of action evolution and scene transitions — when using 2.5 and the user explicitly requests beats / precise timeline, switch to the timeline mode above. Default remains just numbers.)
4. Use 720p, not 1080p — Currently (July 2026), 720p responds best to prompts, fast and cheap. Blindly chasing high definition backfires. (2.5 output tiers officially are 480p / 720p; this rule remains valid.)
5. Clarity via post-production — Upscale to 4K using Topaz after finalizing. (Note: some shots may be over-sharpened.)

Seedance 2.0 Official Supplement (August 2026 official manual; the Simplified School coexists. Only hard constraints and official syntax are added here; full details in references/seedance-2.0-official.md):

• Input Limits: ≤9 images; ≤3 videos, total 2–15s; ≤3 audio files, total ≤15s; Mixed input total limit: 12 files — check asset count before writing the prompt. Generation duration: 4–15s.
• ⚠️ Real Human Face Block: Any image / video asset containing an identifiable real human face will be directly blocked and cannot be generated. Real human texture can only be described in text, not uploaded.
• Only Two Entry Points: First/Last Frame (first frame image + prompt) / Full Reference (multimodal combo). Smart multi-frame / subject reference are not selectable.
• @ Asset Syntax: Multiple @s for the same asset improves accuracy. For multiple subjects, use numbering: "Use the cat in @Video 1 as Subject 1" to prevent confusion. Voice timbre can be specified separately: "Narrator voice reference @Video 1". If no audio asset is uploaded, you can reference the sound within a video.
• Extension Gotcha: When writing "Extend @Video 1 by 5s", set the generation duration to the duration of the added portion (5s), not the total duration.

Seedance 2.5 Official Supplement (August 2026 official manual; writing style remains unchanged. The following are new available spaces and official numbers):

• Duration Options: Single direct generation up to 30s (duration 4–30s). "Ultra-long video" mode up to 180s. Already generated ≤30s videos can be extended multiple times ("matryoshka" style) up to 60s. The 15s/5–6 shot generation strategy remains the same; you can just request 30s when the narrative is longer.
• Input Limits (Multimodal Reference): Up to 30 images (within 4K). Up to 10 video clips, total ≤30s. Up to 10 audio clips, total ≤30s. Supports audio-only generation (2.0 required at least one image or video).
• Official Asset Recommendations: 1–5 reference images work best (6–8 can be tried but require more generation). Subject: 1–5 subjects with single images / multi-views each; for >5 subjects, single-view images are more stable. For multiple views, split into several images showing different views (multiple images, multiple views > single image, multiple views). Subject audio/video: 5–10s works best. 1–8 subjects is optimal. Video editing input clips ≤20s work best.
• The Music Iron Rule is More Solid: 2.5 significantly fixed random subtitles / irrelevant BGM intrusions; it even supports using prompts to remove original video BGM while keeping only vocals (subtitles and other visual elements remain) — the rule "only sound effects, no BGM" remains unchanged, and output is cleaner.
• Promptable Capabilities List (things you can directly control in the prompt under 2.5): Timestamp beats, whitebox control (use untextured 3D whitebox to lock camera position / standing position / motion path), green screen, local removal & editing, spatial perspective modification (change camera position and extrapolate out-of-frame space), seamless two-video transition interpolation, motion transfer (replicate reference video's movement / emotion / camera language / vibe), voice reference, multi-person reference (fix "twin-face" issue), direct writing in minor languages (write directly in Chinese; no translation needed).
• Official Scene Recipes (full details in references/seedance-2.5-official.md; use as needed. For ≤15s ordinary narrative, still default to Simplified School four-segment fill-in):
  • No reference assets / general → Official basic formula: asset description + one-sentence overview + split-timeline plot (forward + reverse description) + global supplement.
  • Need real human face / fix AI face, twin-face → 8-dimension real person formula (includes "preserve real microscopic pores and skin texture" mandatory fidelity suffix).
  • 30s long video → Three modules: multimodal reference layer + global settings (including forbidden items to prevent collapse) + timestamped script & shot breakdown (physical instruction + emotional subtext).
  • 60–180s ultra-long video → Add global parameters at the beginning (duration / aspect ratio); otherwise same as the basic formula, written directly per second segment.
  • Video extension → Must include "extend forward / extend backward / continue / sequel" + connection constraints. If a cut is needed, apply the transition formula (official 11-type transition glossary in the reference). Note: Chinese "extend forward/backward" is ambiguous; first ask "Should the new clip go before or after the original?" The prompt only describes the new segment and declares the original video unchanged.
  • Seamless two-video transition → Begin with "Seamlessly connect @Video 1 and @Video 2, do not modify either original video." Then write Video 1's out point, full-screen transition state, Video 2's in point; lock shape / position / direction / speed.
  • Multi-grid storyboard → Declare "@Image 1 is an N-grid ordered complete storyboard, each grid is one shot." Character / scene use separate @ images. Each grid corresponds to one Shot.
  • Smart editing / advanced editing / video editing → [Marking tool and direction] + [Modification target] + [Action and change] + [Effective timing].
  • Whitebox → Coarse granularity (dynamic skeleton, 5-segment formula) / Fine granularity (render and color, 4-segment formula). Coarse granularity currently yields better results.

simonxxoo's real-world test post: https://x.com/ring_hyacinth/status/2076284324345987386 (relayed by @ring_hyacinth, July 2026). Seedance 2.5 official manual: https://bytedance.larkoffice.com/wiki/RXh5ww6EqighMdkVTMccm2d4n7e (excerpted August 5, 2026: parameters / asset recommendations / capability list incorporated above). Seedance 2.0 official manual: "Jimeng Seedance 2.0 User Manual (New Multimodal Creation Experience)" official docx (excerpted August 5, 2026); key points stored in references/seedance-2.0-official.md. Director workflow (tiers / 9 steps / asset lock table / output modes / timeline modes / label specification): references/seedance-director-workflow.md. Additional director patterns (reference scope contract / character reference range / causal chain / dialogue budget / extension direction / seamless transition / multi-grid / contradiction self-check): references/seedance-direction-patterns.md, adapted from liyue-aigc/seedance-2-5-video-director (September 4, 2026).

References

references/seedance-2.0-official.md references/seedance-2.5-official.md references/seedance-director-workflow.md references/seedance-direction-patterns.md