AI Short Film Prompt Skill (shortfilm-prompt)
Turns any idea into a cinematic, paste-ready video prompt for Sora / Kling / Veo / Seedance, backed by 21 genre templates and per-model compatibility advice.
Pure text skill: only Reads templates/ inside the repo; no network calls, no credential access, no external effects; data flow transparent; IP-name interception warnings plus substitutions; Mx-Shell attribution explicit and MIT covers this skill. Deducted: publisher unverified, and the repo's prompts/ section is All-Rights-Reserved under a NOTICE-based dual-licensing split whose isolation from this skill could not be fully verified statically.
Internally consistent: Steps 1-4, seven hard rules and the 30-second checklist align; template-loading precedence is explicit (SKILL rules win); examples define expected features and one full reference output. Deducted: no execution evidence (static cap 10); per-model duration/filter/negative-field claims (Pika 2.2, Doubao 5s/10s lock) are unverifiable and front-end-dependent; failure feedback on abnormal input relies on default Claude behavior rather than explicit definition.
Very concrete trigger conditions in the description (transformation/multi-shot/emotional narratives); bilingual SKILL (EN+ZH); targets domestically reachable services (Seedance, Kling, Jimeng) — good fit for mainland-China users; Step 1 input gating and template branching are clear. Deducted: no declared non-fit boundaries (e.g., non-video requests), and the 21 template files themselves were not provided, so actual coverage quality cannot be confirmed.
Well layered: SKILL + template library + TESTING + examples; README documents four install paths; MIT license explicit; NOTICE discloses the All-Rights-Reserved restriction on Mx-Shell's original prompts. Deducted: no version number or changelog for the skill; maintenance responsibility and update path only implied; evals.yml references evals/run_evals.py which is not among the provided evidence, so its existence cannot be confirmed.
Core task (paste-ready video-model prompts) is fully designed with clear output format and per-model routing; examples/02 shows a structurally complete sample, indicating clear marginal value over naive prompting. Deducted: static cap 7; only one author-self-assessed complete sample; real-world usability (reroll success rates) unverified by third parties; README tweet stats (2026) cannot be checked.
Multiple evidence types: TESTING.md gives a reproducible A/B protocol with a 10-item checklist, examples state expected features and failure modes, and evals.yml shows CI-eval intent. Deducted: static cap 5; all key evidence (self-assessed sample, model compatibility table, PJ Ace tweet and stats) is author-authored with no committed third-party execution results.
- The skill depends on 20+ template files under templates/ that were not provided in this review; actual template quality and the 21-genre coverage are unverified.
- Per-model duration limits, negative-prompt support, and IP-filter behavior change over time (the docs themselves admit front-end inconsistencies); verify against current platform behavior before use.
- evals.yml references evals/run_evals.py, which is not among the provided evidence; whether CI evals actually run is statically unknown; TESTING.md is a manual protocol only.
- The prompts/ section (Mx-Shell's original prompts) is All Rights Reserved; commercial use requires contacting the original author — only the skill itself is MIT.
- Publisher is unverified by the FollowSkills registry (treat as unknown); the example sample is author-generated and self-assessed — pilot-test outputs at small scale before batch use.
What does this skill do, and when should you use it?
This is a Claude Code Skill that converts an idea into a structured, model-ready AI video prompt. Its methodology is distilled from creator Mx-Shell's workflow — his AI short *Zombie Scavenger* was publicly praised by Hollywood director PJ Ace — packaged by jnMetaCode from roughly 130,000 characters of source material. It uses a five-stage prompt structure with 21 genre templates (transformations, multi-shot narratives, emotional pet/family pieces, commercials, food ASMR and more), seven hard rules, and a pre-delivery self-check. It also ships model-specific guidance for Seedance, Veo, Kling, Sora, Runway, Pika, Hailuo and Wan covering duration ceilings, negative-prompt handling and IP filters.
When triggered, the skill runs a four-step workflow: it checks whether the user already supplied video type, duration, subject, scene and style, asking at most 2-3 clarifying questions otherwise; it loads the matching genre template from templates/ via the Read tool; it writes a complete copy-paste prompt following the five stages (core theme, character & scene, atmosphere & quality, camera rules, storyboard); and it closes with 2-3 sentences on its writing choices plus one line of target-model advice (e.g. Seedance blocks IP names, Runway has no negative prompts). Throughout it enforces seven hard rules: concrete nouns over vague praise, mandatory real camera + lens names, the exact breath-like-float handheld sentence, 'Sound: No score. Production audio only.', at least two imperfection descriptions per subject, restrained endings without FX pile-ups, and IP-name avoidance.
- An AI video creator wants Sora or Seedance to produce a 15-second robot transformation and needs one complete paste-ready prompt from a vague idea
- A short-drama or MV maker planning a 3+ shot edit uses the project-planner template to lock subject consistency and color grade before shot 1, preventing drift by shots 3-4
- A creator making an emotional pet/family/farewell short needs season-plus-light time marking and restrained empty-shot storytelling structure
- A team producing product ads, car commercials or food ASMR clips wants beat-driven or sensory close-up worked examples instead of starting from scratch
- A user working across several video models needs to know how to rephrase one prompt for Kling's banned-word filter, Runway's no-negatives rule, or Sora's three-layer moderation
What are this skill's strengths and limitations?
- Evidence-backed methodology: distilled from the complete workflow behind *Zombie Scavenger*, the film PJ Ace publicly praised, and eval-tested
- 21 genre templates cover transformations, narrative, emotional pieces, ads, ASMR, trailers, micro-drama and more, most with full worked examples
- Concrete per-model guidance for 8 engines (duration ceilings, negative-prompt phrasing, IP filters, language preference) saves real trial-and-error
- Seven hard rules plus a 30-second checklist keep output quality consistent
- Bilingual documentation and templates throughout (README, cheatsheet, FAQ, methodology all have .zh siblings)
- SKILL.md relies on Read and AskUserQuestion tools, so non-Claude Code clients need light adaptation
- The planner-first flow for multi-shot work adds friction if you just want one quick prompt
- Duration/negative-prompt figures are approximations; some (e.g. Pika 2.2) are flagged by the author as unverified, and Seedance/Hailuo data leans on third-party sources
- Mx-Shell's original prompts remain his copyright (ARR) — commercial use requires contacting him; only jnMetaCode's templates and skill are MIT
- Best narrative consistency depends on reference-image uploads and subject locking; text-only results vary by model
How do you install this skill?
Four options: 1) One-line plugin install — run /plugin marketplace add jnMetaCode/ai-shortfilm-prompts then /plugin install ai-shortfilm-prompts@ai-shortfilm-prompts in Claude Code; 2) Clone the repo and run claude --plugin-dir . from its directory; 3) Manual copy — mkdir -p ~/.claude/skills && cp -r ai-shortfilm-prompts/skills/shortfilm-prompt ~/.claude/skills/; 4) As a submodule: git submodule add https://github.com/jnMetaCode/ai-shortfilm-prompts.git vendor/ai-shortfilm-prompts and load with --plugin-dir. No-install users can browse or build prompts at prompts.aiolaola.com.
How do you use this skill?
After installing, trigger it in Claude Code with a slash command, e.g.: /ai-shortfilm-prompts:shortfilm-prompt Write a 15-second robot transformation prompt, green palette, energy core in the belt buckle, post-apocalyptic jungle background. The skill asks up to 2-3 questions if details are missing, then outputs one complete prompt with brief explanations. For revisions, point at a single section — it rewrites only that section.
How does this skill compare with similar options?
The same author maintains a family of skills in the same SKILL.md format (superpowers-zh, agency-agents-zh, openshorts, etc.); this is the video-prompt entry and stacks freely with the others. The README also includes a direct ❌/✅ comparison against the common 'just write keywords' prompting approach.