skill
universal-video-skills
Vendor-neutral AI video production skills for Muse, Claude, Grok, Gemini and other Agent Skills clients: prompt craft, cinematic direction, storyboarding, character consistency, dialogue and audio, as
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/svenkatreddy/universal-video-skills itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 21, 2026Universal Video Skills
A skill pack for AI agents that make video. Nine skills covering the whole job: writing prompts, directing the camera, planning multi-shot stories, keeping faces consistent, handling dialogue and audio, stitching the final cut, checking the result before it ships, and an end-to-end Veo web workflow for when Google's Flow is the backend.
The one thing it refuses to do: lock you into a single video model. Write the prompt once, then translate it for whichever generator you're using (Muse, Gemini/Veo, Sora, Kling, Runway, Pika, Luma, or Hailuo) with a per-model adapter sheet. It works in any agent that reads the open Agent Skills format: Claude Code, Muse, Gemini CLI, Grok, Cursor, Codex, and the rest.
Why this exists
Every video skill I found was welded to one tool: one assumed Muse's built-in generator, another assumed an unreleased Meta model, a third assumed a Chinese API toolchain, a fourth was motion-graphics code with no AI generation at all. The craft inside them was often excellent: prompt grammars, continuity systems, QA gates. But you couldn't take any of it to a different model without rewriting it yourself.
So I took the craft and left the lock-in. The prompting grammar, the continuity discipline, the review gates, the assembly scripts were all rewritten from scratch, in plain English, with the model-specific bits isolated in adapter sheets that are dated and honest about going stale.
The nine skills
| Skill | What it does |
|---|---|
video-prompt-craft | The core: shot-brief grammar, JSON prompting, negative prompts, failure-mode hardening |
cinematic-direction | Camera language, lighting and color, style lenses: directing vocabulary |
story-and-shots | Beats, shot lists, storyboards, review gates, fast-track for simple jobs |
character-consistency | Asset registry, reference sheets, verbatim descriptions: no face drift |
dialogue-and-audio | Voiceover writing and casting, subtitles, mixing |
assembly-and-delivery | ffmpeg stitching, transitions, titles, watermarks, platform exports (16:9/9:16/1:1) |
model-adapters | Per-model translation sheets: Muse, Veo, Sora, Kling, Runway, Pika, Luma, Hailuo |
video-qa | Per-shot checklists, frame evidence, continuity across joins, written verdicts |
muse-google-veo-web | End-to-end Veo via Google Flow on the web: setup, prompting, browser ops, full-duration QA, stitching |
Each skill is a thin SKILL.md that routes into deep reference files.
Install all nine or just the ones you need. They're self-contained and
cross-link with relative paths.
Quick start
npx skills add svenkatreddy/universal-video-skills
Or via npm:
npm install universal-video-skills
Per-agent install paths (Claude Code, Muse, Gemini CLI, Grok, Cursor…) are in docs/install.md.
Then look at the examples: a finished shot-brief, its JSON version, a storyboard, and an asset registry from a small sample project. They're the fastest way to see how the pieces fit.
The workflow, in one paragraph
Plan the story in beats, shot-list every beat, lock your characters and style before generating anything, write each shot as a timed brief with a negative block, translate the brief through the target model's adapter, generate, review every clip against its brief with frame evidence, stitch with the concat script, mix the audio, subtitle everything, export per platform. The skills walk you through each step; the gates tell you when to move on.
Honest limits
This pack doesn't generate video. It teaches agents to direct it. It can't make a model do what the model can't do (readable text, perfect lip sync, and complex physics are still hard everywhere). Model specs go stale, so every adapter is dated: verify before you plan around a number. The full list is in docs/honest-limits.md. Read it before you decide this is for you.
Acknowledgments
The craft here stands on work others published first. The prompt grammar and stitch discipline owe a debt to Mason Mosen's muse-video-prompt-skill; the thin-router architecture, phase gates, and asset-registry ideas to Jiang Yong Luo's muse-video-skill; the shot-brief structure and case standards to ImagineVid's awesome list; the motion grammar and evidence gates to Pluviobyte's video-production-skills; and the vendor-neutral, multi-agent positioning to the wider community (visual-skills, DirectorSKILL, the cinematic-video-prompt-skill, and others). The I2V prompt-subtraction principle comes from 0xzgbot's hermes-media-skill-pack (MIT); the failure taxonomy and layered-gate structure from the indie-qa work in xiaowli's openmontage (AGPLv3: concepts only, no code taken); the embedding-based identity QA pattern from ndpvt-web's capy-video-gen-skill (MIT). Everything here is rewritten in my own words. Borrow the ideas, credit the people.
Contributing
New adapters, durable craft references, and real shipped examples are welcome. Vendors' proprietary features, copied prose, and AI-sounding filler are not. See CONTRIBUTING.md.
License
MIT. See LICENSE.