Make the action clear.

Writing fight and armed-clash prompts

This page brings together official guidance, creators’ public observations, and successes and failures in the archive. The final section, “Adapted examples — not tested,” contains my own rewrites. I have not run them yet. Do not treat them as collected, tested prompts.

1. General principles shared by the official guides

• One moment per segment: Veo recommends focusing a clip on one action beat. Splitting a fight into preparation → contact → reaction is more reliable than describing the whole fight in one sentence.

• Formula: subject + action + setting + camera + style/light. This is Veo’s five-part structure; Wan uses a similar formula.

• Timecodes and shot blocks: Veo supports timecodes such as [00:00-00:02]; Wan 2.6 can use “Shot 1 [0-3s]”; Kling 3.0 supports up to 6 shots in 15 seconds. Shot blocks are also common in the Seedance community.

• Put camera language first: Runway recommends starting with “[Camera] shot of …”. Hailuo accepts camera instructions in brackets, such as [Tracking shot].

• One camera move per shot: Grok Imagine recommends one movement at a time; stacking too many can cause confusion.

• Midjourney Video: for image-to-video, add --motion high for larger movements, --raw for less stylization, and --video 1.

2. Model strengths, based on community tests and collected examples

3. Common failures and how to address them

1. Weapons jump between people or hands: state what each person holds, for example “A holds a bamboo pole, B holds a sickle,” and maintain that assignment throughout.

2. Too many cuts and drifting identities: Seedance creators recommend fewer cuts. Slower, continuous movement is less likely to break down than frantic exchanges.

3. A requested single take still cuts: Langby2 observed edits despite requesting one take. Add “single continuous take, no cuts” and reduce changes of setting.

4. Literal interpretations: Langby2’s crane instruction produced an actual crane in frame; arrows also often travel the wrong way. Standard English camera terms such as “crane shot” are safer.

5. A punch becomes another action: use specific verbs such as jab, uppercut or sweep kick, and describe the reaction after contact.

6. Concealing AI artifacts: godswayfoundinc suggests action-cam handheld shake to make intersecting hands and feet less noticeable.

7. Real-time speed: rather than “no slow motion,” try a shutter description such as “1/500s shutter.”

8. Clothing or faces change: Seedance 1.0 changed clothing in a test. Image-to-video with references is more stable than text alone.

9. More text can hurt: several Seedance users report better results from concise prompts.

4. Violence moderation and policy-compliant descriptions

• Veo (Google): violations return RAI codes, including Violence 61493863 / 56562880 and Celebrity 29310472 / 15236754. These indicate that violence or real-person rules stopped the request.

• Kling: its usage policy prohibits glorifying violence and content involving controlled weapons.

• Seedance: when the level of gore is unspecified, filters may soften the whole sequence. Explicitly stating “no blood, no gore” is often more consistent.

• Policy-compliant approaches:

  • Replace bleeding with sparks, black particles and flying dust.

  • Convey force through sound and reactions — staggering, falling, dropping a weapon — rather than showing wounds.

  • Frame the scene as stylized, fantasy or historical reenactment: stylized, historical reenactment, mythological.

  • Avoid real people’s faces and celebrity names.

  • For burning buildings, describe firelight, smoke and fleeing crowds, not people burning.

• Some community posts share ways to defeat moderation, such as 7724bart’s post. I do not recommend following them: they may violate platform terms and lead to account bans.

5. Adapted examples — not tested

These examples of Taiwanese Hokkien–Hakka armed clashes are my own rewrites based on the points above: adapted examples, not tested. They are not collected prompts with verified results; their performance is unknown.

Example A: a village clash at night (Seedance/Kling shot blocks)

Shot 1 [0-4s]: Wide shot, night, a 19th-century Taiwanese farming village with red-brick sanheyuan houses. Two clans face each other across a muddy field, torches raised. Clan A holds bamboo poles and hoes, clan B holds sickles and wooden shields. Firelight flickers on their faces.
Shot 2 [4-9s]: Handheld tracking shot following a young man in a dark cotton tunic as he blocks a sickle with his bamboo pole and sweeps his opponent's legs. Sparks and dust, no blood. Every character keeps the same weapon throughout.
Shot 3 [9-15s]: Slow crane up over the field, a thatched roof catches fire in the background, villagers scatter, smoke rises into the moonlit sky. Cinematic, realistic, 1/500s shutter, historical reenactment.

Example B: Indigenous spirits in combat (single take)

Single continuous take, no cuts. A mountain spirit made of mossy stone and glowing ferns rises from a misty forest in central Taiwan and clashes with a giant serpent formed of river water. The camera orbits slowly around them at ground level as their impacts throw spray and leaves into the air. Mythological, stylized, dawn light through the fog, no gore.

Example C: hand-to-hand kung fu (image-to-video)

[Tracking shot] Two martial artists in a temple courtyard of an old Taiwanese town. The man in grey blocks a straight punch, counters with an elbow strike, then a spinning sweep kick; his opponent stumbles back and rolls to his feet. Fluid, continuous choreography, real-time speed, no slow motion.

Wan 3.0: official guidance and writing for live action and emotion

These notes are based on official guidance and creators’ public posts. Citations link to their sources; this is not the complete official documentation.

The official prompt guide and Wan 3.0 writing notes

Official guide: Wan 3.0 video generation prompt guide

• Official formula: prompt = overall description + references (image1/video1/audio1) + shot N (start–end seconds): subject + setting + movement + aesthetic controls + dialogue (character says: “…”) + sound effects/BGM + style/mood + negative list.

• Time each shot: use continuous timestamps, with 2–5 seconds per shot (the generation documentation suggests 4–6 seconds). Name transitions such as hard cut or dissolve. For one shot, explicitly request “single continuous take / generate a single shot.”

• Reference media in order: count images, videos and audio separately. The first reference image in input.media is “image1”; the first video is “video1.”

• Dialogue: use character + speaking verb + colon + quotation marks. Add “voice timbre reference audio1” and lip sync.

• Slow motion: write “(1000fps slow motion).” Avoid high-frequency movements such as shaking a head 3 times in one second.

• Official optimization tool: enter /wan3-pe followed by the prompt in the Bailian AI chat to optimize it with the Skill.

• Editing and extension: state the intent with words such as edit video, remove, replace, change to, extend, continue, extend forward/backward. Set ratio to adaptive and duration to -1.

Community test notes (summaries of creators’ own observations, not official guidance):

• Ukiyo_il (original post): assign one character to each reference. Structure 30 seconds as introduction → attack/defense → taking a hit → counterattack → resolution. Describe fast movement as preparation → omitted middle → afterimage → arrival. Specify camera position, rotation direction and endpoint; attacks need origin → contact → result. Keep dialogue short.

• Langby2 (original post): Seedance prompts cannot simply be transferred to Wan 3.0; characters drift even with the same reference. The reported cost for 720P, 30 seconds was HKD16.47, compared with HKD25.6 for Seedance 2.5.

• hii_roto (original post): detailed fight actions improved movement; excessive effects instructions made it stiff.

• The official @Alibaba_Wan account recommends testing at 360p/480p before increasing resolution.

• Reported weaknesses across creators: fast action is less stable than slow scenes, faces drift with multiple image references, and speech sometimes ignores the requested language.

Writing for a live-action look

These points combine the official guide with highly liked live-action examples collected on this page, with sources for each point.

• Light: name a source, not “perfect lighting.” The official formula groups light with shot size and camera movement under aesthetic control (official prompt guide). Popular examples specify direction and temperature, such as a key light 45 degrees above-right at 5600K (johnAGI168). Night scenes use “realistic night exposure, no fake glow effects” (poruru_ai). Explicitly exclude “overly perfect lighting” too (Inshrah_ali_).

• Camera: use real filming language. Official examples name focal length and depth of field, such as an 85mm telephoto with shallow depth of field, with 2–5 seconds per shot. Imperfections can help: slight handheld shake, autofocus and exposure shifts (codewithhajra); an observed rather than performed feeling and imperfect framing (mightyking); 8mm gate weave and grain (old_pgmrs_will). For a premium ad look, use ARRI Alexa color and shallow depth of field (johnAGI168).

• Skin: repeat realistic skin texture in each segment. Describe a real photographic texture, no plastic look and no AI sheen (johnAGI168 IN RED). Stray hair, barely visible stubble and slightly tired eyes (mightyking) are more concrete and less wax-like.

• Expression: describe physical reactions, not adjectives. Instead of “very sad,” specify continuing tears, red swollen eyes, a reddened nose and realistic facial muscle twitching; exclude a staged performance (PixelAigc). Breaking into laughter and a natural ending can also help (MonetizationDon). Break movement down to one core action every 3–5 seconds instead of overloading a second (Tencent News / Wu’an AI test; the official guide also cautions against frequent repeated movements).

• Avoid a plastic look: add exclusions at the start or end. Examples include “100% like real footage, not CGI or animation; avoid overly perfect light, plastic skin, fake backgrounds and oversharpening” (Inshrah_ali_); “practical-looking effects… Avoid an obvious AI look” (ahmadmalik375); no wig-like hair, plastic skin, puffy under-eye ridges or waxy CG faces (johnAGI168); no beauty smoothing (PixelAigc).

• Avoid facial drift or swaps: ① Start with an identity lock listing face shape, skin tone, hairline and stubble; forbid changes of face or age (Xaroon_x), or use separate REFERENCE LOCK and OUTFIT LOCK sections (poruru_ai). ② Use one character per reference (Ukiyo_il); for three people, specify seats and left/right relationships (drjoetw). ③ Follow the official negative list: no facial deformation, face swapping or extra fingers (official prompt guide). ④ Lock the first-frame background for image-to-video (PrometheanAIX). ⑤ Multiple references and fast motion are less stable, so use fewer images and slower action for live-action scenes, testing at 480p first.

Writing emotional scenes

Based on the Alibaba Cloud Wan 3.0 prompt guide, Bailian’s Video Generation Prompt Guide and community-tested approaches. Each point identifies its source and quoted meaning; it does not invent official advice.

Official recommendations — check the originals

1. Include a style/mood section in the formula Source: Wan 3.0 video generation prompt guide Quoted formula, translated: “prompt = [overall description] + … + [style/mood] + [negative list].” Mood examples include tragic isolation, light healing warmth, oppressive tension, relaxed confidence and mysterious fantasy.

2. State the overall mood in the overall description Source: the same guide, shots and timestamps section. Translated quotation: “Overall description: use one sentence to convey the theme, viewpoint, narrative style and overall mood.” Its short-drama example concerns giving up and finding hope again, with glistening tears, loss and helplessness.

3. Voice = words + emotion + intonation + speaking speed + timbre + accent Source: Bailian Video Generation Prompt Guide, sound formula. This is the translated voice formula. Its example gives actionable direction such as a relaxed tone and moderate speed, rather than only “happy.”

4. Match microexpressions to the emotion of the dialogue (audio-driven example) Source: Wan 3.0 prompt guide, audio-driven example. Translated quotation: speaking includes natural blinking, small nods and breathing, with facial microexpressions matching the care and reminiscence in the dialogue.

5. The negative list can exclude exaggerated blinking or gestures Source: Wan 3.0 prompt guide, negative list / action. It says not to exaggerate body movements or blinking. Add these exclusions when you want a restrained performance.

6. The prompt dictionary has a character-emotion field Source: Video Generation Prompt Guide, character emotions. Translated examples describe being very happy with pure happiness shining in the eyes, or a surprised facial expression. They still specify visible eye or facial details, not just an emotion word.

Community-tested emotional direction, with sources

• Write a visible action chain, not just an emotion label Source: Atlas Cloud: writing Wan 3.0 prompts The point is to describe a sequence of observable actions: calling a character anxious states what you want; chewing slowing down, brows tightening and the gaze dropping to the hands tells the model what to show. (English version on the same site)

• Give a long take one slow, visible change of gaze or posture Source: the same Atlas Cloud explanation of official examples — a fixed close-up can specify a verifiable small movement, such as shifting the gaze from ahead down to one’s hands.

MiniMax H3: official writing notes

These notes are based on the official Skill documents. Citations identify the sources; this is not the full official documentation.

Official prompt-writing guidance

Based on MiniMax’s official h3-prompt-writing Skill and base-en.txt, with a source for each point.

1. Choose the mode before applying its structure Source: SKILL.md. T2VA/I2VA/FL2VA/L2VA use three fields: integrated_multimodal_description, overall_soundscape, non_diegetic_music. Ref2VA uses six sections: subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, non_diegetic_music.

2. Match the duration to 4–15 seconds Source: Tips in the same Skill. “Always match the total duration of the description to the requested video length (4–15 seconds).”

3. Prefer concrete visual and audible details to abstract adjectives Source: Tips in the same Skill. Prefer concrete visual and audio details over abstract words like “cinematic” or “beautiful.”

4. Shot timestamps: no time on the first shot, strictly increasing times after it Source: base-en.txt, §4.2. Example: [Shot 2] At 00:03.500, the camera cuts to...

5. Camera movement = type + amplitude + speed Source: base-en.txt, §4.3. For example: The camera pushes in with small amplitude at slow speed toward.... Medium amplitude and normal speed may be omitted.

6. Keep speaker IDs stable and dialogue in its original language Source: base-en.txt, §4.4, and the Skill’s Output Rules. Use (S1)/(S2). Preserve the original language of dialogue, lyrics and on-screen text; do not translate them.

7. Separate the soundscape from the score Source: base-en.txt, §4.6–4.7. overall_soundscape describes environment, action and nonverbal vocal sounds. non_diegetic_music is only music heard by the audience, not the characters. If there is no music, write N/A.

8. Explain how the first and last frames connect to the timeline Source: Skill Tips. “When using keyframes (I2VA / FL2VA / L2VA), clearly state how the first and/or last frame connects to the timeline.”

9. Keep reference labels consistent throughout Source: Skill Tips. Use the same <Picture 1>/<Video 1>/<Audio 1> labels in every section; do not rename them partway through.

10. Atlas and community supplements — useful, but not official wording Source: AtlasCloudAI awesome-minimax-h3-prompts README. Its seven points cover reference binding, observable action, spatial relations, camera and editing, visual style, audio and necessary constraints. BeatAPI’s community archive lists 550+ verified examples: https://github.com/BeatAPI/awesome-minimax-h3-prompts.