菜单
Watch a video that Manifold has pre-staged. Manifold's main process downloads the video, extracts auto-scaled frames with ffmpeg, and produces a timestamped transcript (native captions or gpt-4o-transcribe). This skill receives the report path and tells Claude to read each frame.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use whenever running tests in this project — verifying a fix, validating a PR, running the suite after changes, or checking a single file. Covers which command to use, how the `better-sqlite3` ABI rebuild fits in, and how to target individual files.
Detect documentation pages whose covered code has drifted and open one doc-sync PR per stale page. Use when running the documentation wiki's self-firing lint (scheduled or on demand), or when asked to sync the architecture docs with the code.
Create GitHub issues from agent findings, bugs, regressions, UX problems, or follow-up work. Use when the user asks to create, file, or draft a GitHub issue and the result should be a structured issue body plus a `gh issue create` command.
Sweep the GitHub issue queue and deliver one reviewable PR per labeled issue. Use when running the autonomous issue pipeline's self-firing sweep (scheduled or on demand), or when asked to process issues labeled `llm-github-issue`.
Create GitHub pull requests for completed fixes or features. Use when the user asks to prepare, push, or open a PR and the result should include a meaningful remote branch name, commit, and PR title tied to the actual bug fix or main change rather than an agent or worktree name.
Manifold design system — philosophy, principles, and patterns for UI work. Use when creating or modifying UI components, styling, theming, layout, or any visual aspect of the application.
| name | watch |
| description | Watch a video that Manifold has pre-staged. Manifold's main process downloads the video, extracts auto-scaled frames with ffmpeg, and produces a timestamped transcript (native captions or gpt-4o-transcribe). This skill receives the report path and tells Claude to read each frame. |
| argument-hint | <workdir-path-or-url> [question] |
| allowed-tools | Bash, Read |
| license | MIT |
| user-invocable | true |
The Manifold desktop app (this skill is bundled with it) does all the heavy lifting in its main process. When the user clicks Run in the Watch panel, Manifold:
yt-dlp (or resolves the local path)ffmpeggpt-4o-transcribe (OpenAI or Azure)report.md plus frames/frame_*.jpg into a temp working directory/watch:watch <workdir> into the active Claude Code agentYour job in this skill is only to:
report.md (path is the first argument)cat "$1/report.md"
If $1 does not look like a directory containing report.md, the user invoked
/watch:watch <url> directly without the Watch panel. In that case, respond:
Direct CLI invocation is not supported in this build. Open Manifold's Watch panel, paste the URL, and click Run. Manifold will download the video and hand the prepared frames back to me.
The report lists frame paths under "## Frames" with t=MM:SS timestamps. Use
the Read tool on every frame path in a single message (parallel tool calls)
so you see them together. The transcript section gives you the spoken content
aligned to the same timeline.
If the user asked a specific question (everything after the workdir path), answer it directly, citing timestamps. If no question was asked, summarize the video — structure, key moments, notable visuals, spoken content.
The report ends with a Work dir: ... line. Manifold owns the directory's
lifecycle; do not delete it yourself.