Use grok(grok build CLI) + ffmpeg to accomplish six things:Web/X search, large-batch research, image generation, video generation, green-screen/video post-processing, image post-processing. Images and videos go through grok logged-in built-in tools（image_gen/image_to_video), consuming subscription quota,without incurring xAI API costs。
⭐ Operating principles (read this first)
This skill is guidance, not a hardcoded set of scripts. On each invocation, do this in order:
- First probe the real environment：
node {baseDir}/scripts/preflight.mjs(cross-platform, report OS / grok login / which filters and encoders ffmpeg has / what's missing and should be installed). - If dependencies are missing, install them for the current platform— preflight will give install commands for that OS (install what's needed, e.g. if
drawtextis missing, install a build with freetype; iflibvpx-vp9is missing, switch builds).Don't work around requirements by settling for less. - Media generationUse the wrapper script (
grok-media.mjs, node, cross-platform);ffmpeg post-processingRefer to post-processing.md 's principles and templates,write commands on the spot and tune parameters on the spot based on the material(such as similarity/blend for keying), rather than applying fixed parameters.
You (Claude / grok) can already read the environment, install tools, and write commands — please use your judgment and improvise, treating the scripts below asreliable wrappers / reference implementations, not the only path.
Platform support
| Platform | grok CLI | Description |
|---|---|---|
| macOS / Linux | ✅ | curl -fsSL https://x.ai/cli/install.sh | bash |
| Windows | ✅ Native PowerShell (irm https://x.ai/cli/install.ps1 | iex) or WSL2 |
node scripts are cross-platform; shell commands use PowerShell syntax |
- Media generation requires login (SuperGrok / X Premium+):
grok login。 - ffmpeg post-processing is cross-platform, and the main command body is the same; for complex orchestration, prefer using node to call ffmpeg (avoiding shell escaping differences).
Capability routing
| What you want to do | What to use | Detailed reference |
|---|---|---|
| Web / X search (with citations) | grok-search.mjs(requires key) or grok headless (logged-in state) |
search-and-research.md |
| Large-batch / multi-topic research | grok -p headless + subagents (enabled by default) |
search-and-research.md |
| Generate images | grok-media.mjs image |
media-generation.md |
| Generate videos (text-to-video/image-to-video) | grok-media.mjs video |
media-generation.md |
| Green-screen keying / compositing / watermarking / GIF | ffmpeg (probe → install per platform → write on the spot) | post-processing.md |
Common commands (media generation · cross-platform node)
BASE={baseDir}
# 0. 永远先自检环境
node "$BASE/scripts/preflight.mjs"
# 搜索（JSON-first + 引用；需 XAI_API_KEY）
node "$BASE/scripts/grok-search.mjs" "谁赢得了最新一届世界杯" --mode auto
# 生成图片（登录态，零 API 费；aspect-ratio: 1:1 2:3 3:2 16:9 9:16 …）
node "$BASE/scripts/grok-media.mjs" image --prompt "赛博朋克城市夜景，霓虹倒影" --out city.jpg --aspect-ratio 16:9
# 图生视频（duration 1-15s；resolution 480p/720p/1080p）
node "$BASE/scripts/grok-media.mjs" video --prompt "镜头缓慢推进" --image city.jpg --out city.mp4 --duration 6 --resolution 720p
# 文生视频（自动分镜 → 源图 → 动画）
node "$BASE/scripts/grok-media.mjs" video --prompt "一只猫在爵士酒吧弹钢琴" --out cat.mp4 --duration 6
No fixed scripts for post-processing— first preflight.mjs check the current ffmpeg state, then follow post-processing.md 's templates and write based on the material. For example, the core of green-screen keying is: first have grok generate solid chroma-key green (#00FF00) background material, then ffmpeg -vf "chromakey=0x00FF00:<sim>:<blend>,despill", and tune similarity/blend based on the result.
Mechanisms you must know
- Media generation = prompt-driven built-in tools。
grok-media.mjsuses a strict prompt to callgrok -p "…" --always-approve, having the agent callimage_gen/image_to_videosave to--out, and verify it was written to disk and return JSON. You can also directlygrok -p "用 image_gen 生成…存为 x.jpg"to bypass the script. - Artifact output location: grok first writes to
~/.grok/sessions/<cwd>/<id>/(the video is invideos/), thencpto the working directory; the script is already packaged. - The video comes with an AAC audio track, H.264, 24fps, and the aspect ratio follows the source image / request.
- Time required: about 30–70s for a single image, several minutes for a single video segment; for batch tasks use subagents or a playbook (see examples/video-reel-playbook.md）。
Script list (all cross-platform · node)
scripts/preflight.mjs— environment/capability self-check, reports current status + gives installation suggestions by platformscripts/grok-media.mjs— image/video generation (driven by grok login state, JSON output, cross-platform grok path resolution)scripts/grok-search.mjs— search (xAI Responses API,XAI_API_KEY）scripts/search-selftest.mjs— search script self-test
ffmpeg post-processingdeliberately does not provide a fixed script—it is environment-dependent and parameters need to be adjusted according to the material, see post-processing.md let the agent write it on the spot.
Installation
# 推荐：npx skills（装到 .agents/skills 并 symlink 到各 agent）
npx skills add Wangnov/grok-skills # 所有检测到的 agent
npx skills add Wangnov/grok-skills -g -a claude-code # 全局 + 指定 Claude Code
Both grok build and Claude Code load from ~/.agents/skills/ and both recognize the Anthropic Skills format.