What does Cinematic Continuity Prompt do?
Translate creative ideas into executable, verifiable directorial instructions. Default to Chinese natural language; precision comes from concrete visual facts, causal actions, physical feedback, audio-visual timing, and clear final states, not from vague adjectives or platform format stacking.
First Determine Task and Precision
Input Relationship
First, identify the closest input mode:
- No-Reference Generation: Build a complete audio-visual timeline from text.
- First Frame Driven: The reference image is at second 0. Develop forward using "first frame anchor → action start → continuous development → result or reaction".
- First-Last Frame Driven: Two images lock the start and end respectively. Focus on writing the observable intermediate path and the process of gradual convergence.
- Last Frame Driven: Start from a reasonable preceding state and gradually fall into the composition and state specified by the reference image.
- Multimodal Reference: Images, videos, and audio respectively assume the responsibilities of identity, scene, style, action, camera movement, composition, voice line, or key frame.
- Reference-First Workflow: First design a first frame image suitable for continued motion, then rewrite the video prompt based on the image actually generated and confirmed by the user.
- Prompt Review / Failure Repair: Identify the earliest causal error, conflict, or overload point, then rewrite a complete, copyable version.
Then choose the shooting organization: continuous one-shot, action-matching storyboard, reference image evolution, native phone camera, micro performance, adult restrained intimate interaction. Do not treat one-shot, FPV, aerial shots, or complex camera moves as inherently superior.
Precision Levels
- Normal Mode: Automatically select parameters that truly affect the image. Focus on writing visible results.
- High-Precision Mode: Enable when the user says "detailed prompt", "as detailed as possible", "complete prompt", "complete director's cut", "director-level" or similar expressions. Long prompts are allowed, but each sentence must correspond to a visual fact, continuity requirement, or real failure risk.
If high-precision mode lacks photographic settings, first ask in plain language about the desired effect (e.g., spacious or oppressive, intimate or distant characters, blurred or clear background, handheld or stable, documentary or polished); then translate the answers into equipment type, focal length, aperture, depth of field, focus, and motion parameters. Do not require the user to understand professional parameters first. If the subject is already clearly a phone, drone, surveillance camera, or documentary, directly follow the corresponding device logic.
Use a Three-Layer Architecture
1. Static Visual Layer
Select fields that truly affect generation from the following elements; do not mechanically fill a form:
- Time: Era, season, moment of day.
- Subject: Identity, age, appearance, body, clothing, material, props, initial posture.
- Scene: Natural environment, architectural structure, weather, furnishings, spatial relationships, signs of life.
- Photography: Style, aspect ratio, composition, camera position, angle, perspective, focus, depth of field, light source, light direction, color.
Only write camera brand, specific focal length, and aperture if they can explain perspective, depth of field, or imaging quality and do not conflict with the device mode.
2. Dynamic Continuity Layer
Write each key action as a complete state chain:
Initial state → Visible trigger → Preparation → Start → Process → Body/prop/cloth/environment feedback → Deceleration or stop → Verifiable final state
The character's center of gravity, gaze, hands, prop states, and camera path must share the same spatial logic. Do not repeat completed actions. Objects cannot appear, disappear, change hands, or change material without cause.
3. Audio-Visual Synchronization Layer
Handle ambient noise, action sounds, non-verbal character sounds, dialogue/voice-over, and background music separately. For each key sound, describe its source, distance, triggering action, start and end, intensity, and spatial variation. Do not unconditionally add background music.
See references/prompt-framework.md for the complete template and checklist.
Lock Invariants and Reference Responsibilities
Before writing, determine:
- Identity or structural invariants;
- Initial space, composition, and light direction;
- Camera start, path, and end;
- Subject action and psychological causality;
- Visible details brought by distance changes;
- Wind, cloth, hair, shadows, reflections, and sound feedback;
- Verifiable final frame and interface to the next shot.
For multimodal references, establish a responsibility table. Each piece of material only assumes clear responsibilities; it may have multiple responsibilities but must not vaguely request "reference everything":
| Material | Assignable Responsibilities |
|---|---|
| Image | Identity, clothing, product appearance, scene, composition, style, first frame, key frame, last frame |
| Video | Action, camera movement, editing rhythm, temporal structure, source video editing or continuation |
| Audio | Original sound reuse, voice line, speaking style, rhythm, music style, sound continuity |
Distinguish between "must fully retain", "allow local variation", "only transfer a certain attribute", "only serve as broad atmospheric reference". This is an internal checking language; it is not mandatory to output MiniMax H3-specific fields or English tags.
Resolve Conflicts
Check identity, age, facial features, hairstyle, clothing, time, space, light direction, equipment, depth of field, camera movement, left/right hand, action sequence, dialogue timing, and final state.
Priority:
User-specified reference image facts > User's current explicit instructions > Global unified setting > Repeated and consistent description > Timeline local description > Atmospheric rhetoric
Actively unify when you can judge, and briefly explain in the director's judgment; only ask if two versions are equally critical and would change the final output. The final copyable prompt must not contain mutually exclusive settings.
Organize Timeline and Shots
First organize actions by semantic beats, then map to seconds when the user provides duration, edit points, dialogue synchronization, or platform requirements. For each beat, write clearly: current space, action preparation, action process, feedback, end point, and next trigger.
Camera movement must be reducible to:
Movement type + Amplitude + Speed + Movement target + Stop state
Distinguish between primary camera motion and small human/vehicle perturbations. Push-ins, lateral drifts, and height undulations can be merged into a single motion curve, but cannot be parallel instructions of equal intensity. Prefer continuous camera moves for slight changes in shot size or angle; cuts must bring new space, new time, new information, new state, or a necessary action match.
Distance, Performance, and Physical Feedback
- Long shot: Write scale, direction, and spatial relationships.
- Medium shot: Write posture, center of gravity, clothing silhouette, and character relationships.
- Close-up: Write eye movement, blinking, breathing, mouth corners, contact points, and cloth stress.
- Extreme close-up: Only write pores, vellus hair, powder texture, fine imperfections, glass reflections, or metal scratches.
Expressions must be triggered by an event: Stimulus → Perception → Eye/breath changes first → Head or body reaction → Confirmation → Emotional subsidence → Final posture.
Contact must first establish the contact point and force, then write the response from the other person or object, center of gravity change, cloth change, and stop state. For complex machinery, also write locking, force, motion, deceleration, and stable stop.
Reference-First Workflow
When the user asks to create a first frame or reference image first, read the "Reference-First" section of references/prompt-framework.md and execute:
- Output a concise director's judgment;
- Output an independent, complete, copyable image prompt;
- Make the first frame an open state where action is about to happen or just begins, not a closed poster-like pose after action completion;
- Reserve space for hands, feet, props, gaze, motion direction, foreground-background parallax, and camera path;
- After the user provides the actual generated image, use the visible facts in the image as the new baseline, establish an image-video consistency table, and rewrite the video prompt.
Do not pretend to have locked the specific composition before the user provides the final first frame.
Specialized Modes
- For native phone, POV, micro performance, adult restrained intimate interaction, or dialogue tasks, read references/precision-writing-guide.md.
- For learning user-provided excellent samples or needing similar language density, read references/exemplar-library.md.
- For understanding the trade-offs between the MediaStorm framework, existing methods, and MiniMax H3 official methods, read references/method-comparison.md.
- For updating this Skill or verifying new rules, read references/regression-suite.md.
Output Contract
Unless the user explicitly requests a simplified version, deliver by default:
- Director's Judgment: Core event, visual focus, action causality, camera choice, maximum generation risk; keep concise.
- Full Directly Copyable Prompt: Primarily in Chinese, written in playback order, complete; no "rest unchanged", "same as above", "refer to previous text", or ellipses replacing the main body.
- Stable Retry Version: Only provided when there are too many characters, dense action, multi-hand hand-offs, complex camera moves, multiple reference materials, or other high-risk combinations; reduce concurrent actions without changing the narrative conclusion.
Within the prompt, you may use titles and paragraphs based on the task, but do not force all platforms to use MiniMax H3's three-part, six-part, <Subject N>, <Picture N> or <d> syntax. Only convert to H3 official format when the user explicitly requests it.
Detailed Prohibitions
Prohibitions can be detailed, but must correspond item by item to the real risks of the current shot. First write sufficient positive states, then supplement by category:
- Identity, appearance, clothing, and structural drift;
- Limb, finger, foot, contact point, and force errors;
- Sudden changes in prop quantity, material, position, and state;
- Repeated actions, reversed order, no trigger, no end point, or multiple characters moving in sync stiffness;
- Camera jumps, incorrect zoom, causeless reversal, automatic cuts, and random shake;
- Softening, duplication, or clipping of buildings, machinery, vehicles, and spatial structures;
- Light direction, shadow, reflection, weather, and environment continuity errors;
- Audio-visual misalignment, lip-sync anomalies, sound advance, delay, or repetition;
- Style contamination, wrong depth of field, excessive skin smoothing, commercial retouching, CG or 3D look;
- Garbled text, subtitles, logos, QR codes, and watermarks.
Length is allowed, but synonym repetition and generic negative words unrelated to the current shot are prohibited.
Pre-Completion Self-Check
- Are the three layers consistent and all serving the same core event?
- Is the responsibility of each reference material clear and unambiguous?
- Does the first frame have room for movement? Is there an observable path between the first and last frame?
- Is the camera an executable curve, and are its parameters consistent with the device mode?
- Does every action have a trigger, preparation, feedback, stop, and final state?
- Do changes in shot size bring the correct level of detail?
- Are dialogue, action sounds, and ambient sounds synchronized with distance and the frame they occur in?
- Is the final frame clear, and can it naturally transition into the next shot?
- Are the prohibited items precise, clearly categorized, and not replacing positive descriptions?