Tools· Media Production

claude-video

claude-video is a video analysis pipeline that downloads videos, extracts frames, and produces transcripts so Claude can answer targeted questions about individual scenes.

analyticalstructuringexecutableOpen Source

Source code

In Claude Code: /plugin marketplace add bradautomates/claude-video, then /plugin install watch@claude-video. Other environments: npx skills add bradautomates/claude-video -g. Requirement: yt-dlp and ffmpeg, optionally an API key for Groq or OpenAI as the Whisper fallback.

Open provider

claude-video

Description

The repository provides a skill with the /watch command. A video, whether passed as a URL or as a local file, is downloaded with yt-dlp, ffmpeg automatically extracts scaled frames from it, and a transcript is drawn primarily from existing subtitles. If subtitles are missing, a Whisper API from Groq or OpenAI steps in as a fallback. Automatic frame budgeting ensures that even longer videos can be analyzed without exceeding the token budget.

The result is a searchable context made of imagery and text that makes it possible to answer questions about specific moments in a video, for example what is visible in a particular frame or what was said at a particular point. Locally, yt-dlp and ffmpeg are required; the Whisper fallback transcription additionally requires an API key from Groq or OpenAI.

Strengths

Broad platform support

Through yt-dlp, URLs from YouTube, Loom, TikTok, X, Instagram and several hundred other sources can be processed, in addition to local files in mp4, mov, mkv and webm.

Captions first

At the transcript detail level, yt-dlp returns existing subtitles without downloading the video, covering the majority of public videos for free and instantly.

Frame deduplication

Comparing each frame against the last one kept drops near-identical shots, for example held slides or static screen recordings, so the frame budget is spent on genuinely different content.

Automatic frame budgeting

An auto-fps logic spreads the frame budget across the video length, with denser sampling in focused time windows, so even a long video does not blow the context window.

Multiple host environments

Besides Claude Code, the skill can also be installed via the Agent Skills CLI into Codex, Cursor, Copilot, Gemini CLI and over 50 other hosts.

Assessment

AI features

Multimodal frame reading Claude reads the extracted JPEGs in parallel as images and answers questions based on what is actually visible in the frame, rather than relying on the title or description.
Whisper transcription as fallback If subtitles are missing, a mono audio track is extracted and sent to Groq's whisper-large-v3 or OpenAI's whisper-1, and the backend can also be forced via an option.
Timestamp-driven frame selection If the transcript names a specific moment, for example a cue like look here, the script grabs a targeted frame at that point in addition to the regular sampling.
Scene-based frame detection In the balanced and token-burner detail levels, the video is decoded for scene changes, from which evenly distributed frames are then selected.

Suitable for

Users who want to evaluate video content without watching the entire material manually
Analysis of individual frames or scenes, for example to check image composition, on-screen text or product placement
Transcription of videos without existing subtitles via the Whisper fallback
Teams who want to embed video reviews or content analyses into an existing Claude Code workflow

Limitations and notes

Local dependencies yt-dlp and ffmpeg must be present; on macOS, setup installs them automatically via brew, while Linux and Windows only print the required commands, which then have to be run manually.
API key only when needed Whisper transcription via Groq or OpenAI only kicks in when a video has no subtitles, which mainly affects local files, TikToks, some Vimeo uploads and the occasional caption-less YouTube video.
Frames dominate token cost In a measured 49-minute test video, frames at the default width cost roughly 197 tokens per image, and a resolution of 1024 pixels roughly quadruples that figure.
Warning at token-burner For videos with many scene changes, 116 cuts in the measured test case, the token-burner mode returns every frame uncapped and triggers a warning past 250 frames, while balanced thins the result to 100 frames.

Quick start

Choose the installation path for your environment: in Claude Code with /plugin marketplace add bradautomates/claude-video, then /plugin install watch@claude-video, on other hosts with npx skills add bradautomates/claude-video -g.
On the first /watch call, the script automatically checks whether yt-dlp and ffmpeg are present and sets them up via brew on macOS.
Pass a video and a question, for example /watch https://youtu.be/dQw4w9WgXcQ what happens at the 30 second mark.
Focus on a time window with --start and --end if needed, which raises the frame density in that window.
Optionally set an API key for the Whisper fallback, GROQ_API_KEY is preferred, OPENAI_API_KEY as an alternative, in ~/.config/watch/.env.

Tips

Use --detail transcript when only the spoken content matters, which skips the frame download entirely.
For on-screen text, such as slides, terminals or code, set --resolution 1024 so the text stays readable.
Use --no-dedup only when near-identical frames should deliberately be kept; normally deduplication saves tokens without losing information.
For long or high-motion videos, try --detail efficient first, which often delivers enough frames in a fraction of the time.

Access

The repository is public at https://github.com/bradautomates/claude-video and is released under the MIT license. In Claude Code, installation runs through the plugin marketplace: /plugin marketplace add bradautomates/claude-video, then /plugin install watch@claude-video. For other environments such as Codex or Cursor: npx skills add bradautomates/claude-video -g. For the web version, the skill file is available for download and is added via Settings, Capabilities, Skills. yt-dlp and ffmpeg are required, plus optionally an API key for Groq or OpenAI as the Whisper fallback.

Alternatives

descript.com — freemium

Descript

Commercial tool for transcription, video editing and scene search, an alternative to your own pipeline.

otter.ai — freemium

Otter.ai

Commercial transcription service offering similar question-and-answer functions for recordings.

rev.com — kommerziell

Rev

Professional commercial transcription and subtitling service, a replacement for the Whisper pipeline.

Last reviewed: 27 July 2026 · Pricing, plans and features are a snapshot in time. Check the provider's own page before deciding.

In the workshop this becomes your method.

Whoever sees this process run once wants the agent behind it next. We build that in the workshop From Process to Agent.

View workshops

Turns call notes or a sales call transcript into a CRM-ready follow-up document with needs analysis, BANT assessment and a scheduled follow-up plan.

Learn more

Conversation, not pitch

Understand first, then decide. We take time for an initial conversation, without sales pressure, without obligation.

Schedule a call