Design action fight scene prompts for AI video generation (miora and other channels). Goals: cinematic camera movement, clear rhythm, believable impact, and no teleporting, clipping, or ineffective punches.
Four Iron Rules (all are indispensable)
Before designing any fight segment, first build the skeleton according to these four rules, then add the flesh:
- Temporal segmentation: Split the total duration (≤15 seconds) into multiple small stages of 2-3 seconds each. Each stage is responsible only forone group of main actions, explicitly marked with timestamps (such as
0:00-0:02). Do not cram a dozen simultaneous actions into one prompt. - Action causal chain: The actions of adjacent stages must have a physical causal relationship—opponent charges forward → the female lead uses the momentum to turn; the female lead hits the abdomen → the opponent's body leans forward → only then can the next stage grab the shoulder. Every action must be able to answer "what conditions did the previous action give it?" Isolated actions = the root cause of teleporting/clipping.
- Environmental feedback: Each stage must include at least one instance of environmental feedback (dust flying, metal clashing, sand and rocks collapsing, cracks in the ground, debris splashing). This is not decoration—it tells the model where the force comes from, and the sense of impact is grounded by this.
- Asset locking: At the beginning of the prompt, use a separate paragraph to declare the character's face, hairstyle, clothing, weapons, and body proportions asunchanged throughout. In two-character scenes, you must separately state which character each reference image corresponds to (
图1 = 女主，图2 = 怪兽), to prevent structural confusion.
Advanced Rules (determine the upper limit of the final film's quality)
The four iron rules ensure it is "watchable"; the following rules ensure it is "good-looking":
- Sound effects as an independent layer: Write the sound effects section separately—either add a short sound effect sentence at the end of each stage, or write a full "sound effect rhythm" summary at the end. When no music is needed, explicitly write "no music, keep only sound effects/ambient sound." Sound effects must cover three layers: force (engine revving higher, dull thuds), material (gravel, friction), and creature (roars, breathing).
- Action results may fail: A hit ≠ a successful hit wins. A blade cannot cut into hard armor and is deflected, the recoil sends the protagonist flying, the blade edge curls—physically consistent frustration is more credible than crushing the enemy the whole time, and it also provides a narrative hook for tactical shifts.
- Trajectory-decomposition style description: For a key strike, do not write the result; write the trajectory—"falling from the air → both hands raised high → vertical heavy slash → the blade edge lands exactly on the thick armor." The four-step chain makes it executable frame by frame for the model.
- Give specific durations for slight slow motion: For the slow motion at the moment of impact, do not write "slight slow motion"; write "about 0.2-0.3 seconds of slight slow motion, then immediately return to normal speed." At most once per stage.
- The ending must set up the next round: The afterglow stage is not just a freeze frame—the vehicle's front points toward the target again, realizing the armor cannot be broken, tearing off the work badge and walking toward the exit. Leave a hook for the next segment.
- Write group enemies in layers: In multi-enemy scenes, specify the number and the order of actions ("6 hunting companion monsters all turn around and roar, among which 2 are the first to pounce"), and the shot must clearly capture that order; in group battles, multiple opponent faces must all be different, and must not resemble the protagonist.
- Style switch upfront: Overall style declarations such as one continuous shot, handheld phone documentary texture, no stabilization shake, etc. are placed at the very front of the prompt, declared only once, with no mid-course switching.
Output style (choose one of two, depending on subject matter)
- A. Timestamp style：
0:00-0:02 段落正文. Suitable for fast-paced close-quarters fighting and car chases—dense rhythm, frequent shot changes. In the main text, camera language is naturally embedded in the sentence pattern "the camera uses ... to follow and shoot." - B. Shot-naming style：
镜头01——接近. Suitable for large scenes and major jumps in shot size (tracking against ocean waves → orbiting → extremely distant aerial shot); each shot gets a verb-based name, and a single shot size is fully written out within the paragraph.
Both styles follow the four iron rules, and the same prompt must not mix them.
Workflow
Step 1: Collect input (ask for all missing items at once)
- Total duration (miora H3 channel is an integer from 4-15 seconds; do not submit outside the range, explain to the user)
- Characters: how many sides? Human/monster? Each one's weapon? Any vehicles?
- Reference images: how many are there, and which one corresponds to whom
- Scene environment: ground material (sand/stone/metal/ruins/water surface)—determines the environmental feedback vocabulary
- Subject type: close-quarters fighting / vehicle chase / group battle / giant creature showdown—determines which scene recipe to use (see
references/scene-patterns.md) and the output style - Sound effect requirements: whether music is needed (by default no music, only sound effects)
Step 2: Build the rhythm skeleton
Allocate the stages according to the "three-beat pattern," then fine-tune:
| Beat | Proportion | Content |
|---|---|---|
| Charge-up | ~30% | Confrontation, closing in, first probing clash |
| Outburst | ~50% | Main clash: continuous chain of cause-and-effect actions, with the densest strikes |
| Afterglow | ~20% | Ending freeze frame: victory and defeat are clear, dust settles, camera slowly pushes in |
Output a stage table: fill in 4 columns for each stage—timestamp / shot / main action (one group) / environmental feedback. Show this table to the user first, and after confirmation, expand it into the complete prompt.
Step 3: Write the complete prompt
Assemble in this order (optional sections are used as needed):
【风格开关段】一镜到底/手持纪实质感等整体声明（如有）
【资产锁定段】人物外观声明 + 参考图映射；群战加对手面孔规则
【场景段】环境、光线、色调、氛围粒子（尘埃/水雾/粉笔灰）
【时序段】逐阶段：时间戳或镜头命名 + 镜头语言 + 动作 + 因果衔接词 + 环境反馈
【音效段】分段音效或整体"音效节奏"总述；默认声明"不要音乐"
【技术段】画幅、帧率、画质参数（16:9、24 FPS 等按项目规范）
For the vocabulary and examples of the four iron rules, see references/camera-and-impact.md; for the four subject-matter recipes and the skeletons of the two output styles, see references/scene-patterns.md; for a complete example (female lead vs monster, 15 seconds), see references/example-15s-duel.md。
Step 4: QA check
Before delivery, self-check item by item, and revise anything that does not pass:
- Each stage has ≤ one group of main actions, with a duration of 2-3 seconds
- Any two adjacent stages can be connected by cause and effect ("because A, therefore B" holds)
- Each stage has ≥ 1 instance of environmental feedback
- There is an asset-locking section at the beginning; in dual-character scenes, the reference image mapping is clear; the rules for opponent faces in group battles are complete
- Camera movement varies (not a fixed camera position throughout), and the camera serves the action (close follow/handheld shake during the outburst stage, slow push or freeze frame during the afterglow stage)
- No isolated actions, no "teleport-style" displacement description
- Key strikes include a breakdown of the action trajectory; the moment of impact is slowed down with a specific duration (≤ once)
- Has an independent sound effect layer, and the "no music" constraint is specified as required
- The ending has setup for the next round (turning to the target / tactical shift / emotional freeze-frame)
- The output style is consistent throughout (timestamp style or shot-naming style, not mixed)
Boundaries
- Only design the prompt text, do not submit generation tasks (generation goes through the corresponding skills such as miora-video-studio).
- When the duration exceeds the channel limit, do not split and submit—whether to split into shots and batches is decided by the user.