Watch CLI

Watch any social video → get an architecture diagram, working component, runnable notebook, or step-by-step cheat sheet — automatically.

watch-cli is a thin orchestrator that downloads any social video, extracts evenly-spaced frames, and transcribes the audio. The output is a single labeled block (or one-line JSON) designed for an LLM to read frames as images and transcript as text. The agent supplies the prompt; watch-cli supplies the raw materials.

When to invoke

Reach for watch whenever the user gives you a video URL and wants you to do something with what's in it.

• The user pastes a video URL with no verb. Ask one clarifying question ("summarize, implement, clone the UI, extract the architecture, or something else?"), then run watch.
• The user asks to "summarize", "explain", or "walk me through" content at a video URL.
• The user asks to "implement", "clone", "build", or "replicate" what is on screen in a video.
• The user asks to "extract architecture from", "diagram", or "turn this paper talk into code" at a video URL.

Supported platforms: YouTube, X / Twitter, LinkedIn, TikTok, Vimeo, Reddit, Facebook. Every URL is fetched anonymously. A login-walled URL fails with tag=download-auth; watch-cli does not touch browser sessions on its own. If the user wants to use one, they opt in per run with WATCH_BROWSER=auto (or a browser name), which lets yt-dlp read cookies from the local browser profile and send them only to that platform; --cookies <file> uses an exported cookie file instead. Ask the user before setting either; never set them silently.

What you get back

The watch output gives you the raw materials to map to five concrete artifacts. Match the user's intent to one of them.

A coding walkthrough becomes working project files: read the frames for file names, exact code, and dependencies; read the transcript for intent and rationale; emit the final state, not the intermediate edits.

A system architecture talk becomes an interactive architecture diagram: a single self-contained HTML page with actors, surfaces, APIs, and clickable named flows that highlight the path through the system.

A UI or motion demo becomes a working React component: one paste-ready .tsx file that captures the feel and the single interaction that makes the UI special, not a pixel-perfect screenshot.

A paper or research talk becomes a runnable notebook: one .ipynb implementing the core method on a toy dataset that runs end-to-end on a free Colab T4.

A long tutorial becomes a step-by-step cheat sheet: numbered steps with copy-pasteable commands, video timestamps, verification per step, and only the troubleshooting the speaker actually discussed.

Five copy-paste prompt templates live in prompts/. Pick the one that matches the user's intent.

Parse rules

watch emits a versioned, agent-shaped payload. Both formats are documented in docs/output-schema.md and conform to the v1 contract — append-only, no renames, no type changes within v1.

• Preferred: JSON mode. watch <url> --format json emits one UTF-8 JSON object on stdout terminated by a newline. Parse it with a real JSON parser, switch on obj.version, read obj.video_path, obj.duration_sec, obj.frame_paths (array of absolute JPG paths, earliest-in-video first), obj.transcript (string or null), and obj.exit_code. Field reference is the single source of truth in docs/output-schema.md.
• Fallback: text mode. Some agent hosts (Claude Code does this today) capture stdout as a free-text block. The leading line WATCH_OUTPUT_VERSION: 1 is the version signal; everything below is labeled blocks (VIDEO:, DURATION:, FRAMES:, TRANSCRIPT:, EXIT:) per the same doc.
• Read frames as images, transcript as text. Each path under FRAMES: (or each string in frame_paths) is an absolute path to a JPG on disk. Pass the path to the host's image-reading primitive. The transcript is plain UTF-8 text — no decoding needed.
• Exit-code behavior is documented in docs/exit-codes.md. The partial-success case is the one to remember: on exit 4 the frames are populated and the transcript is null — branch on the exit code and fall through to a frames-only consumption path instead of failing the run.

Invocation

watch <url> [frame-count]

Default frame count is 8. For a fast-cut or dense UI demo, double it. For a multi-hour conference talk, bump to 24–32. The CLI does not cap; agent hosts typically prefer ≤ 32.

Check the archive before watching

Every successful run is stored under ~/.watch-cli/archive. Re-watching a URL is a cache hit — no download, no transcription, and the output block is byte-identical to the cold run — so re-running a URL is cheap, but re-deriving an answer already on disk is wasted work.

watch-archive find "context graph"   # → id, [04:32], the matching line
watch-archive ls                     # what has already been watched
watch-archive get <id|url>           # reprint one record, transcript timestamped

Reach for watch-archive find first when the question is "have I already seen something about X?" or "where in that video did they say Y?". It searches every stored transcript and answers with a timestamp, which is a seek position rather than a video to sit through again.

Records are plain JSON, SRT and JPG on disk — grep and jq read them without this CLI, and <id>/transcript.srt loads in any video player. Layout in docs/archive.md.

Pass --no-cache only when the source itself has changed. A failed transcription is never stored, so a retry after an error always makes a real attempt.

Anti-patterns

• Do not parse stderr. Progress lines on stderr ([watch] downloading …) are not part of the contract and change between releases. Programmatic consumers ignore stderr.
• Do not parse the filename of a frame to infer its position. Filenames are implementation detail; the ordering of frame_paths is the contract.
• Do not hard-fail on every non-zero exit. exit 4 is recoverable partial success — frames populated, transcript null. Branch on exit_code before parsing.
• Do not surface API keys or environment variable names in chat. KYMA_API_KEY setup lives in the README.
• Do not embed the locked pitch into a longer marketing paragraph. The description line above is the source of truth; reuse it verbatim where the host shows skill metadata.

What leaves the machine

• The video download goes to the platform hosting it, through yt-dlp.
• The extracted audio track is uploaded to Kyma API for transcription; frames and the video file stay on disk under ~/.watch-cli/archive. With --with-local installed, transcription runs offline through whisper.cpp instead and nothing is uploaded.
• Browser cookies are never read unless the user sets WATCH_BROWSER or passes --cookies; when they are, they go only to the platform that set them.
• Installation is pinned: the release installer downloads a tagged tarball and verifies its SHA256 against the checksum published on the same GitHub Release. Homebrew does the same through the formula's sha256.

Transcription runs through Kyma API. Get a key at https://kymaapi.com/?src=skill:watch and set KYMA_API_KEY; a one-hour video costs about $0.05.