MiniMax H3 video reverse skill
The skill retells a reference video, waits for your confirmation and writes MiniMax H3 prompts per scene and shot.
- 325 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The skill runs local scripts and reads frames of your video.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/minimax-h3-video-reverseDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The skill runs local scripts and reads frames of your video.
This is third-party code. Review the repository files before installing.
What it does
MiniMax H3 video reverse is a skill that breaks down a reference video and turns it into MiniMax H3 prompts. The agent first retells what it saw: what happens, how actions connect and where the clip ends. You confirm or correct that understanding, and only then does the agent deliver text. There are four modes: `AUTO` chooses the route itself, `T2VA` writes a description of the whole clip for text-to-video, `I2V` gives a first-frame prompt and a motion prompt for each shot, and `DUAL` produces both versions from one fact table. Conversation runs in Russian or Chinese, while the prompt itself is English by default. The scripts `scripts/validate_prompt.py` and `scripts/build_review_sheet.py` check the format and build a frame contact sheet with real timestamps. The skill contains no H3 model and does not generate video itself.
Who it is for. Video makers and prompt engineers who reproduce the structure of a reference clip in MiniMax H3.
Good fit when
- A ten-second reference needs a shot-by-shot first-frame and motion prompt.
- A long clip is split into segments, for example thirty seconds each.
- You want to confirm the agent understood the plot before it writes a prompt.
- One version for text-to-video and another for image-to-video are needed.
Not a fit when
- The agent cannot actually read a whole video.
- You need the video itself: the skill only writes prompts and renders nothing.
- The source video belongs to someone else and no rights to recreate it were obtained.
Example request
Break down the attached video, retell the plot and wait for confirmation. Then give a MiniMax H3 prompt in I2V mode.Limitations
The agent must actually watch the video, otherwise the retelling becomes guesswork; without that ability the skill works only from a text description. `scripts/build_review_sheet.py` needs Pillow installed. The scripts read frames of your clip locally, so check that the file holds no one else's personal data. Rendering happens on the chosen video platform and the skill asks for no API key. You can recreate someone else's clip structure only if you hold the rights. The last commit dates from October 2026.
How to disable. Delete the `minimax-h3-video-reverse` folder from the agent's skills directory, for example `${CODEX_HOME:-$HOME/.codex}/skills/`.
FAQ
Is an API key needed?
No. The skill does not generate video and asks for no key; the clip is created on the platform you choose.
How do I skip the confirmation?
State that confirmation is unnecessary and the prompt can be written right away. By default the agent retells the clip first.
How do T2VA and I2V differ?
`T2VA` gives a single description of the whole timeline for text-to-video. `I2V` gives a first-frame prompt and a motion prompt for each shot.
How do I install it in Codex?
Create `${CODEX_HOME:-$HOME/.codex}/skills` and clone the repository into the `minimax-h3-video-reverse` subfolder. For other agents, copy the repository into their skills directory.
Can a finished prompt be checked by script?
Yes, in a plain chat you can run `scripts/validate_prompt.py`: it catches format violations such as multi-source shots and extra targets.
Discussion
No questions or reviews yet.
Related
SkillsThree Codex and Claude Code skills: search MiniMax H3 cases and full prompts, pick tutorial routes and prepare submissions
A director skill carries a short drama from script to a ready MiniMax H3 prompt package with H1 to H36 checks.
A Claude Code skill that breaks a script into shot lists for MiniMax H3 and teaches facial direction that renders.
The skill turns a short idea into a full MiniMax-H3-Image prompt and can render the picture through the MiniMax API.
Codex skill that makes short MiniMax H3 videos from your images on a Google Colab GPU, with a job queue and result download
Local video with MiniMax H3 on your own GPU: the open-video CLI and a skill that lets an agent write prompts and run clips.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| MiniMax H3 video reverse (this card) | no | 325 | |
| Cinematic Video Prompt Engineer | no | 148 | writes cinematic prompts from scratch; this skill dissects an existing clip and confirms understanding with you |
| Realistic Video Prompting | no | 337 | describes a home-shot look for several models; this one targets MiniMax H3 |
README in short
The skill keeps one rule: understanding first, text second. Versions 0.5 and 0.6 added checks of visible shot details and returning images for visual review.
README badge
Original SKILL.md
--- name: minimax-h3-video-reverse description: 先复述参考视频内容与叙事逻辑，和用户确认理解后，再反推 MiniMax H3 文生视频或逐镜图生视频提示词。Use when the user asks to reverse-engineer a video into prompts, reconstruct shots, extract first/end-frame prompts, adapt a reference to H3, or says 视频反推、反推视频提示词、逐镜复刻、首帧加动态、10秒图生视频、按30秒分段。Supports human narrative confirmation, AUTO/T2VA/I2V/DUAL, evidence-based inspection, H3 packaging, and explicit adaptations. Does not generate videos or recover a hidden original prompt by itself. --- # MiniMax H3 视频反推 先与用户对齐“视频在表达什么、事件为什么相连、目标要保留什么”，再把能够确认的空间、时间、动作与声音转成生成指令。动作描述完整，不代表故事理解正确。 默认分两阶段：**看视频并复述理解 → 用户确认或纠正 → 输出提示词**。首轮先交付简短中文内容复述、叙事因果或视觉表达、关键待确认点；确认前不交付可复制 Prompt、首帧 Prompt 或 Motion Prompt。用户明确跳过确认时遵从并记录跳过；“一段提示词”“只要英文”只是输出格式，不自动跳过。 默认保真反推。用户明确要求压缩、扩写、换主体或改成一镜到底时，保留原片事实，另写目标方案与改编记录。不要把改编当作观察结果。 ## Resource Guide - 开始分析前读 `references/video-observation.md`：能力门控、五遍读取、疑点驱动返图、时间与声音边界。 - 素材含音轨或用户要求继承、反推声音时读 `references/audio-inheritance.md`：原声复制、音频参考、声音重建与音画同步的分流。 - 建立源事实后、写任何提示词前必读 `references/narrative-confirmation.md`：内容复述、因果核对、纠正循环、确认复用与失效。 - 确定交付模式后读 `references/output-contract.md`：AUTO 路由、T2VA、I2V、DUAL 和分段交接。 - 写 H3 可复制块前读 `references/h3-adapter.md`：三字段、首尾帧对齐、能力核实及格式冲突处理。 - 动作、遮挡、变形、混合媒介或改编遇到困难时读 `references/historical-lessons.md`。 - 查来源、版本边界与未验证事项时读 `references/provenance.md`。 - 生成反推生产包前读 `references/evidence-contract.md`：事实绑定、确定性编译、复核失效、双模式一致与分段继承。 - 示例在 `examples/`；v2 输入示例见 `examples/evidence-contract.input.json`；旧 `examples/contract.json` 仅作格式草稿。 ## Workflow ### 1. 接收与能力门控 提取：素材、期望模式、保真/改编、目标时长、语言、平台能力、声音目标。只问会改变任务的缺失信息；未指定模式用 AUTO，未指定语言用英文 Prompt 加简短中文说明。保真任务中若确认素材有音轨而用户未要求换声，默认**保留原始音轨**；这不等于已听懂，也不等于三字段文字 Prompt 能自行复制音频。按 `references/audio-inheritance.md` 选择可执行的继承路线。 必须确认能否访问完整画面、有效视觉时间线和足以辨认事件的连续内容。链接页面、缩略图、抽帧命令成功或文件含音轨，都不等于已看过/听过。 - 只有不可访问链接、零散截图或文字：完整视频反推为 `BLOCKED`，返回缺失输入与继续条件；可交付明确标为“局部画面草稿”的静态描述，不能冒充全片。 - 画面完整、声音无法核实：`PARTIAL`，继续视觉反推；写 `audio_status: unavailable`，不补造声音。 - 画面与用户要求的通道均已核实：`COMPLETE`。有音轨但用户只要求视觉，可标 `not_requested`，说明仅完成视觉范围。 - 不把 `BLOCKED` 输出包装成生成就绪的 H3 Prompt。 如果可用本地 Python 与 FFprobe，可先运行： ```bash python3 scripts/probe_video.py /path/to/reference.mp4 ``` 脚本只核对文件和时间元数据，不观看画面、不听声音，也不下载链接。视觉/听觉检查必须用宿主真实提供的能力；缺少时如实降级。