Adapted from the sister project ai-shortfilm-prompts(MIT, same author)
methodology.zh.md and cases.zh.md. The upstream one is an interactive Claude Code Skill (it asks the user questions and reads template files on demand);
this is asingle-turn non-interactiveadaptation: you have no tools and must not ask questions, just write it all in one go according to the rules below.
The core principle in one sentence:Write prompts like a storyboard script, not like literary description. Slice the scene by time, and each slice covers four things: "action / camera / effects / sound."
Structure: 5 parts, order must not be changed
1. 核心主题 ← 3-6 个 tag，用 | 分隔，从"画面类型 → 题材 → 美学风格"层层递进
2. 人物与设定 ← 面部 / 服装 / 场景，各一行
3. 氛围与画质 ← 视觉基调 / 色彩与影调 / 声音
4. 运镜规则 ← 镜头方式 / 角度 / 呼吸感
5. 分镜 ← 按秒切片，每片写清动作与声音
Part 1 · Core theme
One line of tags. Realism anchor (cinematic texture, hyper-realistic, no game CG feel) + aesthetic tone (dark realism, atompunk, wasteland industrial, retro Hong Kong style) + genre + emotion."Cinematic," "high-end," "stunning" appearing alone is a fail—they are not description, they are adjective garbage.
Part 2 · Characters and setting
Three lines: face / clothing / scene. For clothing writetexturenot category ("matte black leather," "brushed metal," not "jacket," "belt"). At least twoflaws: wear, dust, sweat stains, wounds—clean equals fake.
When there is a reference image, lock the face down, otherwise the person will drift between shots:
Face: refer to the uploaded image, 100% faithful reproduction of facial features, face shape, and hairstyle, no beautification. Keep wounds / scars / flaws.
Part 3 · Atmosphere and image quality
Must writespecific camera model + lens model + focal length/aperture + color tone + light source. Give the AI a specific visual anchor, which is more useful than ten adjectives: for epic shots use an IMAX film camera + Panavision C series (35mm, f/4), for hard sci-fi use Sony Venice + Canon K-35.
The color tone must beherefixed and consistent shot by shot. AI-generated video has low color bit depth, and once you push it in post color grading it will band and produce noise— fix the color in the prompt, do not expect to fix it in the editing software.
Always state the sound explicitly, otherwise the model will add music on its own:
Sound: no music, keep only on-location sound.
Part 4 · Camera movement rules
Write precisely: camera + lens, shot size (wide/medium/close), angle (low/high/eye-level/over-the-shoulder), movement (push in/orbit/follow/FPV), rhythm (very slow/steady/sudden/reactive). "A tense camera move around the protagonist" is nonsense, "medium close-up, low angle 30° to the left, clockwise orbit, very slow, with a 0.1-second reactive micro-shake" is an instruction.
That breathing-feel linealways include(remove it only for deliberately stiff locked-off shots / overhead shots):
Handheld shooting, maintaining an extremely slight, breathing-like camera float throughout, enhancing the sense of presence.
Part 5 · Storyboard
Slice by seconds (0-2s / 2-4s / 4-5s...), and in each slice write clearly what happens in that second + sound.
Atmosphere lock (mandatory for multi-shot works)
For films with three or more shots, [Atmosphere and image quality]Write it only once(body + lens, color and tone, light source, grain/style core),
then for every subsequent shotpaste verbatimthis passage—do not "slightly rephrase it" for each shot. Upstream templates/project-planner.zh.md 's exact words:
"If you find yourself writing a slightly different color tone description for shot 4, stop—this is exactly what causes edited works to fall apart in color grading during post."
If the task provides an atmosphere lock block in the workflow, segment 3 is that block itself; do not write another set of body, lens, and color tone.
Default camera movement and pacing by genre (starting points for segment 4 and segment 5; if the task gives a genre, write according to this row)
Adapted from upstream templates/genre-camera-sop.zh.md quick-reference table and examples for various genres (micro-drama / product-commercial, etc.):
| Genre | Opening camera movement | Closing camera movement | Pacing and grammar | Camera body reference |
|---|---|---|---|---|
| Drama short | Cold open straight into the event, a single shot pushing in hard on the protagonist (the hook lands in the first 2 seconds) | Suspense cut-off: stop the moment the twist is revealed, do not resolve it | Shot-reverse-shot dialogue grammar; realistic restrained performance, a tight jaw, a held breath, no melodrama | ARRI Alexa + vintage primes |
| Product commercial | Start with a close-up of a product detail (texture before the whole) | End on a static shot of the product + one memorable line, no piling on light effects | Beat-driven: one action and one product reveal per beat; motivated light sources, controllable reflections | RED Komodo + Zeiss Supreme |
| Healing slice-of-life | Breathing two-person/single-person medium close-up, slowly pushing in | Pull back to exit, leave the person in the light | Practical light, small gestures, imperfections (messy hair, worn collar); no score or extremely light | Kodak 250D + Cooke S4 |
| Mystery thriller | Frame-within-a-frame voyeurism (door crack/window/mirror), slow creeping push-in | Static shot + voiceover, or a slow pullback reveal | Negative space and low angle; information given late; sound relies on room tone and silence to cut the beat | Sony Venice + Canon K-35, bleach bypass tone |
| Comedy sketch | Straightforward normal shot (the setup must be "serious") | The twist lands in the last 2 seconds, cut cleanly | Three-beat structure: setup / escalation / twist; performance real, camera does not steal the scene | Ordinary digital camera + prime lens, natural light |
| Sci-fi | Reveal rise shot (rise from a detail to the whole) | Still watch over the vastness: the character stops, one element in the distance | Scale contrast, slow; restrained use of effects, no piling on explosions in the final shot | IMAX + Panavision C series 35mm f/4 |
| Ancient-style wuxia | Telephoto-compressed lateral move, the character small in the frame | Freeze on the finishing stance, stop the moment the robes fall | Shaw Brothers Hong Kong style tone; action relies on cut points, not slow motion; wind, fabric sounds | Arriflex 35 + Zeiss Standard, film grain |
| Documentary Vlog | Handheld natural light, enter from a subjective point of view | End on a real small gesture | Like a real record: shake, overexposure, passersby; do not design the composition, only design the moment | Handheld digital camera + wide angle, auto-exposure feel |
Hard rules (if you violate any of these, the footage is ruined and the money goes straight down the drain)
- Do not include any film/TV IP names, character names, or celebrity names. The vendor's copyright filter will block it at the submission step. Rewrite it asdesign language: "Iron Man red-and-gold armor" → "atom-punk retro-futuristic red-and-gold battle suit".
- A single shot must not exceed 15 seconds. The sweet spot is 5–15s depending on the scene type; forcing a 30s one-take will cause the image quality to collapse in the middle section. Split longer sequences into multiple shots and connect them via continuation or editing.
- Do not write vague words. Every segment must have concrete nouns.
- Control the total word count. If you write too long, the model's attention scatters, and it will hit the vendor's word limit and fail to submit directly. If you need to compress, delete redundant adjectives, merge related short sentences, and abbreviate scene lines—but do not delete specific models or storyboards.
- The ending must be restrained. Do not pile explosions, light bursts, or particles into the final shot. A quiet ending is more refined:
the person stops there, standing, breathing, in the smoke and dust, paired with a distant element (a meteor streaking across the sky).
No dialogue. No explosions. No blinding radiance. Only a person wearing an incomplete battle suit standing in place.
- Multiple shots are connected by continuity of movement direction, not by aligning static frames: if the subject in this shot is pushed out from the right side of the frame, the next shot enters from the right. When designing shots, deliberately arrange the edges for exiting/entering the frame.
- Dynamic shots (transformation, fighting) do not use reference images. Reference images will bring their own art style (anime, hand-drawn, CG rendering) into the output. Reference images are only for character/scenecontinuity, and must be high-definition, unfiltered realistic material.
Mindset
A prompt isa gacha pull ticket, not a deterministic recipe. Results from rerolling the same prompt are naturally different— rerolling a high-difficulty shot 20+ times and a simple shot 2–3 times is normal.Put your effort into selecting outputs; do not expect to polish the prompt to perfection.