Agent Skills › hi5jeff/deepclonewebsite › mdbox-media

mdbox-media

GitHub

提供基于mdbox CLI的多媒体生成与处理功能，支持图文音视频创作、背景移除、超分及OCR文字识别，自动管理文件存储与URL输出。

.claude/skills/mdbox-media/SKILL.md hi5jeff/deepclonewebsite

触发场景

用户请求生成图片、视频或音频需要进行图像到图像的转换或背景移除要求对文档或图片进行OCR文字提取请求对媒体内容进行放大或增强

安装

npx skills add hi5jeff/deepclonewebsite --skill mdbox-media -g -y

更多选项

非标准路径

npx skills add https://github.com/hi5jeff/deepclonewebsite/tree/main/.claude/skills/mdbox-media -g -y

不安装直接使用

npx skills use hi5jeff/deepclonewebsite@mdbox-media

指定 Agent (Claude Code)

npx skills add hi5jeff/deepclonewebsite --skill mdbox-media -a claude-code -g -y

安装 repo 全部 skill

npx skills add hi5jeff/deepclonewebsite --all -g -y

预览 repo 内 skill

npx skills add hi5jeff/deepclonewebsite --list

SKILL.md

Frontmatter

{
    "name": "mdbox-media",
    "description": "Generate and process media — images, video, audio, music — and read text from images (OCR). Use whenever the user asks to create \/ draw \/ render an image, make a video, generate audio\/music, do image-to-image \/ image-to-video, remove a background (cutout), upscale, or extract \/ recognize text from an image, scan, or document (作图 \/ 生成图片 \/ 生成视频 \/ 配音 \/ 海报 \/ 识别文字 \/ 提取文字 \/ OCR \/ 扫描).",
    "allowed-tools": "Bash"
}

MDBOX media generation

Generate and process media with the platform mdbox CLI (already installed; uses the org gateway, no API key needed here). Supports text-to-image/video/audio, image-to-image, image-to-video, background cutout, and upscaling.

Where to save — ALWAYS `./generated_images/`

Save every file the user should see/download into ./generated_images/ (run mkdir -p generated_images first). Only this folder is publicly served for the project.

How to deliver — paste the REAL public URL

The project's public base is in the env var $PROJECT_MEDIA_BASE. For a file generated_images/<name>.<ext>, its URL is ${PROJECT_MEDIA_BASE}generated_images/<name>.<ext>. Output it in chat by type so it renders in web and Telegram/Discord:

Image → ![<alt>](<url>.png) — shows inline
Video → [<title>](<url>.mp4) — web embeds a player; TG/Discord show a link
Audio → [<title>](<url>.mp3) — web embeds an audio player
PDF / HTML / Excel / Word / PPT / other → [<filename>](<url>.<ext>) — clickable link

Always give the user the link — never just say "saved to generated_images/".

Generating — ALWAYS use the `media` directive

To generate ANY image/video/audio you MUST emit a __media__ directive — do NOT run mdbox gen for generation (it blocks your whole turn). Emit it on its own line and finish your reply right away; the system submits it, shows a live progress card, and delivers the result into the chat automatically.

{"__media__": {"kind": "image", "model": "nano-banana-pro", "prompt": "a red fox in snow, cinematic"}}

kind: image | video | audio. model: pick from mdbox models. prompt: required.
Image-to-image / image-to-video: first mdbox upload <file> to get its URL, then pass it in params, e.g. {"__media__": {"kind":"image","model":"nano-banana","prompt":"make it watercolor","params":{"metadata":{"image":"<url>"}}}}.
Emit the directive ONCE per file; don't also run mdbox gen for the same thing.

The mdbox CLI (non-generation)

mdbox guide — full reference (commands + current models). The model list lives here; never hardcode model names.
mdbox models — list current models (for the model field above).
mdbox upload <file> — local file → the https URL that reference inputs require.
mdbox gen --model <model> --prompt "<text>" -o generated_images/<name>.png — SYNCHRONOUS (blocking) generation. Use ONLY when you need the file IN THE SAME TURN (e.g. make an image then immediately feed it into a video). Never for a plain request — use __media__.

Reading text from images (OCR)

To extract text FROM an image (scan, screenshot, photo of a document, receipt…), use mdbox ocr — it is SYNCHRONOUS and fast (~1s), so run it directly in your turn (do NOT use __media__, that's only for generation).

mdbox ocr <image>... — one or more local paths or https URLs (1–16 images). Local files are auto-uploaded. Prints the recognized text (layout order).
mdbox ocr <image> --json — full JSON: full_text, per-image pages[], low_confidence_words[] (with bbox), and stats (word count, avg confidence).
--lang ko,zh — optional BCP-47 hint(s); omit to auto-detect (~99 languages).
-o out.txt — write the text to a file instead of stdout.

Example: mdbox ocr ./invoice.png --lang en → prints the invoice text. Then answer the user's question about it, or paste the extracted text back into chat.

See references/models.md for model-selection guidance (read on demand).