🌐 Browse the prompt library online → prompts.aiolaola.com ��� view any template/case and copy it with one click, no install.
📖 免费配套学习 → 从零学会 AI 编程：180 节免费实操课 + 《AI 编程实战三卷书》在线阅读 + 实战社区 · 永久免费
The complete methodology + prompt library + Claude Code Skill behind Zombie Scavenger by Mx-Shell — the AI short that Hollywood director PJ Ace called "one of the best short films I've seen in years."
⚡ Quick start: one-page cheat sheet · failure→fix gallery
"This is one of the best short films I've seen in years. Very soon, we'll stop calling it 'AI film' and just call it film."
Film name: Zombie Scavenger by MX-Shell.
— PJ Ace (@PJaccetturo), May 10, 2026
| 13.4M views | 82K likes | 7.4K reposts | 39K bookmarks | 2.3K replies |
|---|
Stats are from PJ Ace's original tweet (@PJaccetturo, May 10, 2026), as of mid-May 2026.
This repository is the complete workflow behind that film, made available because Mx-Shell himself published his prompt collection documents and walked through his entire method on a public Douyin livestream.
Copy this into Sora / Seedance / Kling / Veo. Replace {{...}}:
Anamorphic widescreen cinematic. Simulated IMAX film camera +
Panavision C-series lens (35mm focal, f/4 aperture). Handheld
shot — extremely subtle, breath-like camera float throughout.
{{your scene description}}.
No score. Production audio only.
Why this works: real camera bodies + "breath-like float" anchor the AI to actual film aesthetics — not the vague "cinematic feel" keyword everyone else uses. Full breakdown in methodology.md.
Same idea: a female mech warrior raises an energy shield in a thunderstorm.
❌ The naive prompt (what most people write):
Epic cinematic shot of a beautiful female mech warrior activating a
stunning energy shield in the rain. Highly detailed, 4K, photorealistic,
movie-quality, dramatic lighting.
Vague praise — epic / stunning / 4K / movie-quality — gives the model nothing concrete to anchor to. You get generic game-CG output.
✅ With the 5-stage method (excerpt):
Core theme: gritty hard sci-fi mech | rainy dock | battle-damage aesthetic | energy shield | post-apocalyptic live-action
Atmosphere: simulated IMAX film camera + Panavision C-series (35mm, f/4). Low-saturation teal base, film grain.
Camera: handheld — extremely subtle, breath-like float throughout.
9–12s: hexagonal energy cells light up unevenly, some flicker as if faulty; rain bends around a 2m dome.
Ending: no dialogue, no light burst — just rain vaporizing on the shield, a lightning flash across the dock.
Real camera/lens names + physical reactions + battle damage + an empty ending = visceral realism. Full sample with the 10-item self-check: examples/02-skill-output-sample.md.
May 2026. A 29-year-old vocational-school graduate from rural Yunnan, China — handle Mx-Shell — used 10 days and ~20,000 RMB of cloud credits to make a 3-minute AI short called Zombie Scavenger: an atomic-punk robot wanders into a beachfront villa after a zombie apocalypse, meets a confused ostrich, and starts dancing 1980s-style breakdance moves while kicking a zombie's head across the floor.
On the cost: the widely-quoted "3,000 RMB" figure came from Mx-Shell himself, but on livestream he later revised it to "tens of thousands / 20k+." The real spend was likely well above 3,000 — still far below a comparable live-action shoot.
Hollywood director PJ Ace (@PJaccetturo) retweeted the film, calling it "one of the best short films I've seen in years" and started a search for the author.
A few weeks later Mx-Shell went on a Douyin livestream and gave away his entire workflow — the prompts, the camera language, the failure modes, the reroll counts.
This repo is the result of digesting 130,000 characters of his materials into a structured, reusable system.
ai-shortfilm-prompts/
├── README.md You're here. English entry point.
├── README.zh.md Chinese version.
├── cheatsheet.md One-page cheat sheet (the whole method at a glance).
├── cheatsheet.zh.md Chinese version.
├── cases.md Failure→fix gallery: common bad output and the fix.
├── cases.zh.md Chinese version.
├── methodology.md The 5-stage prompt structure, explained.
├── methodology.zh.md Chinese version.
├── faq.md Q&A: tools, failures, costs, edge cases.
├── faq.zh.md Chinese version.
├── credits.md Sources & attribution.
├── credits.zh.md Chinese version.
├── showcase.md Things people made with the method.
├── CONTRIBUTING.md Submission template & rules for new prompts.
├── LICENSE MIT (this work)
├── NOTICE Attribution + Mx-Shell ARR details (dual-licensing)
│
├── prompts/ Mx-Shell's complete original prompts.
│ Body kept in Chinese (his authorial
│ voice), with English header on each file.
│ ├── README.md Index of all prompt archives
│ ├── index.json Machine-readable index of every prompt
│ ├── zombie-scavenger.md *Zombie Scavenger*
│ ├── kamen-rider-transformations.md Kamen Rider transformation × 6 (5 riders + flight)
│ ├── kaisa-transformation.md LoL Kai'Sa transformation × 3 versions
│ ├── pacific-rim-gundam.md Pacific Rim + Gundam mech-drop
│ ├── cyber-wuxia.md Shaw Brothers + steampunk wuxia template
│ └── metal-gear-charge-combat.md Weapon-charge + combat composite
│
├── templates/ IP-stripped generalized templates (EN + .zh.md siblings).
│ Authored by jnMetaCode based on Mx-Shell's structure.
│ ├── 15s-transformation.md 15-second transformation
│ ├── multi-shot-narrative.md Multi-shot edited narrative
│ ├── atmosphere-prefabs.md 8 reusable atmosphere/look prefabs
│ ├── negative-prompts.md Reusable negative-prompt prefab (per-model)
│ ├── genre-camera-sop.md Camera-movement SOP for 5 film genres
│ ├── camera-move-library.md 50 camera moves in 5 technique modules
│ ├── pet-lifetime-narrative.md Worked example: pet × family emotional narrative
│ ├── family-recipe-farewell.md Worked example: a mother's recipe handed down
│ ├── elderly-cat-companion.md Worked example: grandmother & cat, empty-nest
│ ├── product-commercial.md Worked example: beat-driven product hero ad
│ ├── food-asmr.md Worked example: food ASMR / native synced audio
│ ├── animal-vlog.md Worked example: talking-animal selfie vlog
│ ├── movie-trailer.md Worked example: escalating teaser trailer
│ ├── cyberpunk-city.md Worked example: atmospheric neon-city roam
│ ├── claymation.md Worked example: 12fps stop-motion / claymation
│ ├── nature-timelapse.md Worked example: landscape timelapse, locked grade
│ ├── found-footage-horror.md Worked example: CCTV / security-cam dread
│ ├── anime-to-real.md Worked example: 2D character → live-action
│ ├── music-video.md Worked example: beat-synced performance MV
│ ├── sports-slowmo.md Worked example: high-speed slow-motion sports
│ ├── fashion-film.md Worked example: editorial fashion film
│ ├── travel-vlog.md Worked example: travel vlog, sense of place
│ ├── drone-fpv.md Worked example: FPV/aerial continuous flight
│ ├── micro-drama.md Worked example: vertical short-drama (hook+cliffhanger)
│ ├── sci-fi-space.md Worked example: hard sci-fi space / zero-G
│ ├── car-commercial.md Worked example: automotive commercial
│ └── dance.md Worked example: continuous-motion dance film
│
├── assets/ Diagrams + the README hero-demo prompt.
│ ├── demo-prompt.md Copy-paste 15s prompt the Skill wrote (hero slot)
│ └── 5-stage-structure.svg The structure diagram
│
├── skills/shortfilm-prompt/ Claude Code Skill
│ ├── SKILL.md How Claude should generate prompts (7 hard rules + 10-item checklist)
│ ├── TESTING.md How to run rigorous skill tests in another Claude window
│ └── examples/ 4 test cases (5 files) with expected outputs
│
└── .claude-plugin/ Plugin metadata (plugin.json + marketplace.json)
Every Mx-Shell video prompt follows the same skeleton. The order matters:
1. Core theme ← 3-6 style tags separated by |
2. Character & scene ← Face / clothing / environment
3. Atmosphere & quality ← Visual base / color tone / style core
4. Camera rules ← Single-shot or multi-shot / angle / breathing
5. Storyboard ← Per-second OR per-shot breakdown
-
Specify real camera + lens models. Don't write "cinematic feel". Write "simulated IMAX film camera, Panavision C-series lens, 35mm focal, f/4 aperture." AI training data binds those exact strings to real movie aesthetics.
-
Describe imperfections. "Battle-damaged armor, paint worn off, oil in the joints, minor facial blemishes preserved." Perfection looks fake. The visceral realism comes from the flaws.
-
Leave the ending empty. "No dialogue. No explosion. No blinding light. Just a figure standing in the smoke, a meteor crossing the sky."
Full methodology in methodology.md.
/plugin marketplace add jnMetaCode/ai-shortfilm-prompts
/plugin install ai-shortfilm-prompts@ai-shortfilm-prompts
Then in Claude Code:
/ai-shortfilm-prompts:shortfilm-prompt Help me write a 15-second prompt for a
robot transformation, green color palette,
energy core in the belt buckle,
post-apocalyptic jungle background
git clone https://github.com/jnMetaCode/ai-shortfilm-prompts.git
cd ai-shortfilm-prompts
claude --plugin-dir . # then: /ai-shortfilm-prompts:shortfilm-promptmkdir -p ~/.claude/skills
cp -r ai-shortfilm-prompts/skills/shortfilm-prompt \
~/.claude/skills/git submodule add https://github.com/jnMetaCode/ai-shortfilm-prompts.git \
vendor/ai-shortfilm-prompts
claude --plugin-dir vendor/ai-shortfilm-prompts # loads the plugin + skillThe Skill walks through the 5-stage structure, runs a 10-item self-check, and warns you about IP names that may be blocked by Seedance 2.0.
The 5-stage structure is model-agnostic. Here's how the major 2026 engines compare — single-shot ceiling, negative-prompt support, IP filter, preferred prompt language, and the one gotcha that trips people up:
| Model | Max single shot | Negative prompt | IP filter | Lang | Notes / gotcha |
|---|---|---|---|---|---|
| Seedance 2.0 (Doubao / Jimeng / Xiaoyunque, ByteDance) — Mx-Shell's primary engine | ~10–15s — but the Doubao app is locked to 5s/10s preset buttons; the full 4–15s range only on Jimeng/Dreamina web + VolcEngine console | Partial — no reliable dedicated field in the consumer app; negate by describing what you do want | Strict — domestic platforms aggressively reject named celebrities + branded IP | Either (ZH native, EN works) | Duration depends entirely on the front-end. Don't promise 15s if the user is on the Doubao app. Native synced audio-video is its standout strength. |
| Veo 3 / 3.1 (Google) | 8s per clip (4/6/8s); Extend adds 7s hops up to ~148s, but quality degrades after 4–5 extensions | Yes — dedicated field. List unwanted elements as plain nouns (extra limbs, glitch morphs); no no/don't command phrasing |
Strict — rejects public figures, brands, voice/likeness; scans prompt and frames | EN | The negative field wants descriptive phrases, not commands — no rain style instructions can backfire. Best-in-class native audio + realism. |
| Kling 2.x / 3.0 (Kuaishou) | 2.5: 5–10s (Pro ~12s); 3.0: up to ~15s single-prompt multi-shot | Yes — dedicated field. Use for stability artifacts (sliding feet, extra fingers, morphing), not generic "quality" words | Strict — a pre-gen banned-word filter rejects the whole prompt on one match; hypersensitive | Either (ZH native, EN strong; 3.0 multilingual dialogue) | The banned-word filter is notoriously over-sensitive — a benign body/contact word can block a clean prompt. Sanitize wording first. Excellent action/motion realism. |
| Hailuo / MiniMax (02 / 2.3) | ~6–10s — 1080p caps ~6s, 768p extends ~10s | Yes — but vendor guidance says use sparingly, for specific artifacts not as a primary lever | Moderate — more permissive than Sora/Veo, still blocks named celebs + overt IP | Either (ZH native, EN solid) | Resolution and duration trade off — you can't get max of both. Pick the axis that matters per shot. Strong motion at low cost. |
| Wan 2.x (Alibaba, open-source) | 2.2: ~3–8s @ 24–30fps; 2.5/2.6 extend ~10–15s by mode | Yes — robust field; defaults like morphing, warping, face deformation, flickering |
Lenient — open-weights/self-hostable, so no enforced filter when run locally (hosted APIs may add their own) | Native ZH (both, but ZH-trained) | Leans Chinese — first/last-frame mode especially; EN-only prompts can underperform, add ZH for tricky shots. Self-hostable, full ComfyUI control, renders legible ZH/EN on-screen text. |
| Runway Gen-4 / 4.5 | 5s or 10s per generation | No — does NOT support negatives. avoid X / no X can produce the opposite. Describe only what should appear |
Strict — blocks celebs, real people, copyrighted characters/brands | EN | Negative prompts are actively harmful here — no distorted hands can summon distorted hands. The single biggest mistake when porting prompts. Director-grade camera control + mature pro pipeline. |
| Pika (2.2 / 2.5) | Standard + Pikascenes: 5s or 10s; Pikaframes (keyframe) up to ~25s | Partial — 2.5 supports negatives (no morphing, no extra limbs); 2.2 unclear, verify in-app |
Moderate — blocks overt celebs/IP, generally more relaxed than Sora/Veo | EN | Only the Pikaframes keyframe path reaches ~25s — ordinary text/image-to-video is still 5s/10s. Fast, effects/transition-driven, great for stylized short-form. |
| Sora 2 / 2 Pro (OpenAI) | Up to ~25s continuous single-pass on Sora 2 Pro (standard tier shorter) | No dedicated field — negate in-prompt with guardrails like original characters only, no logos or text |
Strict — triple-layer moderation; blocks named IP and visual lookalikes even without the name | EN | The filter catches descriptions, not just names — a spiky-haired ninja in an orange jumpsuit still trips it (matches Naruto). Avoid recognizable trait-bundles, not only proper nouns. Leading prompt comprehension + world coherence. |
Durations and negative-prompt mechanics for Veo 3.1, Runway Gen-4, Kling,
Wan, and Sora are consistent across multiple 2026 vendor/help-doc sources;
Seedance 2.0 and Hailuo figures lean on third-party guides (treat ~ as
approximate). "Veo ~148s", "Sora/Pika ~25s" come from extension/keyframe
features, not plain single-shot generation. IP-filter "strictness" labels
are qualitative.
Tool names link to their sign-up pages; Xiaoyunque / Jimeng / Kling are invite links (both sides get free credits).
This is the video-prompt sibling of the AI-coding ecosystem maintained by @jnMetaCode:
- superpowers-zh — Chinese-enhanced edition of
obra/superpowers(TDD / debug / git workflow skills) - agency-agents-zh — 211 plug-and-play AI expert personas
- agency-orchestrator — Multi-agent collaboration orchestrator
- ai-coding-guide — 66 Claude Code tips
- shellward — AI Agent security middleware
- ai-coding-trilogy — AI coding three-volume book
All projects share the same SKILL.md format. The video skill stacks
freely with any of them.
The roadmap is to collect more AI-shortfilm creators' methods, structured the same way. New prompts, templates, fixes, and translations are welcome — see CONTRIBUTING.md for the submission template and rules (public source required, credit the original creator).
Made something with the method? It goes on the showcase.
微信公众号「AI 不止语」 · 抖音 @AI不止语（AIBZY） · X @jnMetaCode · aiOlaOla — 免费学 AI 编程
🌐 Live prompt library → prompts.aiolaola.com
MIT License for everything authored by jnMetaCode (methodology, FAQ, templates, Skill, metadata). Free for any use including commercial — just keep the copyright notice.
Mx-Shell's original prompts and document excerpts — sourced from his fan-group documents and public Douyin livestream that he himself distributed — remain © Mx-Shell, all rights reserved. Archived here for educational reference; commercial use requires contacting Mx-Shell directly. Full dual-licensing details: NOTICE · attribution: credits.md.
"For creation, the equipment is not what matters. The idea is what matters." — Mx-Shell, May 12, 2026 livestream