open-source-video-gen-skill
mini-yifan/open-source-video-gen-skill
open-source-video-gen-skill is an agent skill project by mini-yifan on GitHub: Vibe Video: runs open-source MiniMax H3 on AutoDL to render video; its short-drama template chains brief, episode synopses, script and dialogue, art and video, with H3 prompt-writing guidance. It covers characters, episodes, script, prompts and production, with prompts for MiniMax H3 and ComfyUI (local), and its README says it works with Cursor. 35 stars, last updated 2026-09-11, no license declared.
Depends on AutoDL GPU instances (billed by time); the short-drama template only handles workflow gates and is thin on narrative
Editor's notes
The repo has 9 skills: besides the short-drama template there is a fashion-model video template, character turnaround sheets, music and Qwen3-TTS voice-over, and music and voice-over share the AutoDL instance used for video. The short-drama template tracks each stage in a progress file (制作进度.md) and resumes from the first unapproved stage, producing per episode a script, a subtitle and dialogue timeline, an art set and video clips. It skips expression reference images, does not generate storyboard reference images by default, and test-renders one clip before any batch. Installation symlinks the skills into ZCode or Cursor, needs Python 3.9+, and comes with a doctor.sh self-check script.
- Type
- All-in-one: script → storyboard → video
- Stages
- Characters · Episodes · Script · Prompts · Production
- Video models
- MiniMax H3 · ComfyUI (local)
- Agents
- Cursor
- Format
- Agent skill
- Language
- Multilingual
- GitHub stars
- ★ 35
- Last updated
- 2026-09-11
- License
- None declared
- SKILL.md files
- 9
- Checked by hand
- 2026-10-13
In the author's words
Vibe Video: one-sentence idea → finished film. Open-source MiniMax H3 + AutoDL elastic GPUs + agent-managed power on/off, about ¥0.3 for 10 seconds of 1080p. An AI film crew of 10 skills (screenwriter/art/director/score/voice-over) that can make short dramas, promo videos and creative videos.
Translated from the Chinese original.
Skills in this repo (9)
Names and descriptions come from each SKILL.md frontmatter; Chinese descriptions are shown in our English translation.
- autodl-app-instance
Powers AutoDL Art application instances on and off via official API, finds the ComfyUI panel by probing the instance's proxy service domains from the snapshot API (no SSH needed, works on macOS / Linux / Windows), and waits until ComfyUI is ready for MiniMax H3. Use when the user mentions AutoDL apps…
- beauty-video-gen
Generates hyper-realistic, lifelike, approachable vertical fashion short videos of beautiful women (cloud generation with MiniMax H3). Use when the user asks to "generate a beauty video / realistic beauty / approachable beauty video / beauty short video / fashion beauty video / custom beauty video", or uses style words such as "cute, friendly, sexy, innocent, mature", or body words such as "exaggerated figure / long legs / curvy figure / legs filling most of the frame / low angle to lengthen the legs" to request a video of a real woman; trigger even if the user does not say "beauty", as long as the request is to "generate a short video with a real woman on camera". Integrates prompt writing from the "Real-Person Fashion Short Video Prompt Writing Framework" and optional face reference image generation (only when the user explicitly asks to "lock the face first / make a face reference image"; by default it goes straight to text-to-video; image generation prefers the current agent's built-in…
- character-three-view
Prepares reference art for video production from a script or asset list: character reference sheets (including outfit-state variants), prop references, and scene concept images, then verifies each generated image. Works for any project style — photoreal, anime, cartoon, fantasy creatures. Use when…
- cursor-image-gen
Generate or edit bitmap images by launching a local Cursor Agent (cursor-grok-4.6-high) and using Cursor GenerateImage, then saving files to an exact folder. Use when the running agent has no native image tool, or the user explicitly requests Cursor, for photos, illustrations, icons, banners, mocku…
- h3-prompt-writing
Write MiniMax H3 video generation prompts for T2VA, I2VA, FL2VA, L2VA, and Ref2VA. Use when rewriting multimodal requests into H3 prompt structures, composing integrated_multimodal_description, overall_soundscape, and non_diegetic_music, aligning keyframes, defining reference labels, or when the ca…
- minimax-h3
Use a remote SeetaCloud or AutoDL ComfyUI GPU instance to generate MiniMax H3 videos. Connects over SSH, discovers the current H3 workflows, helps choose U03/U06/U02, uploads references, submits a job, and downloads the result. MiniMax H3 supports 5–15 seconds per generation; choose the duration fr…
- minimax-music-gen
Generates music with MiniMax Music 3 (ComfyUI) on an AutoDL GPU instance: complete songs with sung vocals (custom lyrics, female/male voice, any genre) or instrumental BGM. Use when user wants to generate music, songs, or audio tracks: "generate music", "write a song", "turn lyrics into a song", "generate background music / score / instrumental", "play me a song", "MiniMax Music", "BGM", "backing track", "love song" all trigger; match equivalent phrases in any language. Do NOT use for music pl…
- qwen3-tts
Generates voice-over speech (TTS) with Qwen3-TTS on an AutoDL GPU instance. Use when the user needs narration, voice-over, dubbing, line reading, text-to-speech, a specific voice type (such as a sweet girl voice, a middle-aged man's voice or a mature woman's voice), cloning someone's voice, or adding a dialogue/narration track to a MiniMax H3 video. Trigger words include: "TTS", "dubbing", "narration", "speech synthesis", "text to speech", "timbre", "voice type", "clone a voice", "voiceover", "text-to-speech". Three voice controls: designing a voice from a text description (VoiceDesign), 9 built-in voices + style instructions (CustomVoice), and reference-audio cloning (VoiceClone). Do not use for generating songs or background music (use minimax-music-gen for that…
- short-drama-production
Orchestrate an end-to-end short-drama or short-film production from a user's theme: approved brief, then — for multi-episode projects — a graded episode outline, then per-episode screenplay with dialogue, art reference assets, and video generation with optional music and dubbing, to a finished vide…
Install
npx skills add mini-yifan/open-source-video-gen-skill
# install a single skill
npx skills add mini-yifan/open-source-video-gen-skill --skill autodl-app-instance
# pick agents (claude-code / codex / cursor / opencode / gemini-cli / openclaw)
npx skills add mini-yifan/open-source-video-gen-skill -a claude-code -a codexDefault commands use the skills CLI and were not verified one by one; follow the repo README. Read SKILL.md before installing. Install guide →
Similar projects
- jingju-skills★ 37
Jingju: full AI manju workflow covering outline, character consistency, episode scripts, storyboard, image/video prompts and ComfyUI batch image generation
- ai-manga-drama★ 17
AI manju workshop: eight-step flow from one-line idea to final video with task-card scripts, character reference library, style anchors, storyboard, image/video prompts, voiceover, subtitles and final compositing check
Hardcore Director: film and AI video controller that routes by stage to screenwriting/short drama, characters and scenes, storyboard, Seedance 2.5/Wan 3.0/H3 prompts, sound and editing modules
- ai-short-drama★ 44
Xuefeng payoff drama: expands one line into a full brief, writes episode scripts on a payoff-beat model while syncing character/scene/prop files, then breaks shots into Seedance 2.0 prompts
Short-drama production controller: novel outline/characters/script, director's storyboard, prompts, sound assets, rough cut and QC (16:9 by default)