claude-video¶

Development line: project:claude-video · thread claude-video-development
Last event: 2026-07-27 · 1 dated since 2026-07-27 · Researched: 2026-09-05 · confidence: medium

What it is¶

claude-video is a /watch skill for video analysis for agents.

• Input: takes a URL or a local file.
• Extraction: pulls frames and subtitles.
• Audio fallback: uses Whisper when subtitles are missing.
• Result handoff: passes output to the agent.

The default limit for long videos is up to 100 frames; token-burner mode is unlimited. It fits screen recordings, video clips, and targeted timestamps, but context cost grows with the number of frames.

Development line¶

• 2026-07-27 — claude-video repository publicly referenced. On 2026-07-27, an entry linked to the bradautomates/claude-video GitHub repository. This is the earliest public reference in the project history and provides a concrete source to inspect.

What changed¶

• 2026-07-27 — repository link recorded. Primary sources show no separate commit or release on that exact date.
• 2026-06-29 — version 0.2.0 added four detail modes, frame deduplication, automatic splitting of long audio for Whisper, and a portable skill package for multiple hosts.

How to use this¶

From 2026-07-27, treat the linked bradautomates/claude-video GitHub repository as the source for implementation and usage details.

1. Install the skill for Claude Code through the marketplace, or for Codex and other Agent Skills hosts with npx skills add bradautomates/claude-video -g. — https://github.com/bradautomates/claude-video
2. Run /watch <URL-or-path> <question>. It supports URLs handled by yt-dlp, and local MP4, MOV, MKV, and WebM files. — https://github.com/bradautomates/claude-video
3. Pass --start and --end for a targeted excerpt. This samples frames more densely while spending fewer context tokens. — https://github.com/bradautomates/claude-video
4. Select transcript, efficient, balanced, or token-burner. If subtitles are missing, set a Groq key for whisper-large-v3 or an OpenAI key for whisper-1. — https://github.com/bradautomates/claude-video

Best practices¶

• Narrow the time window with --start and --end for questions about a specific moment, instead of scanning the full video. — https://github.com/bradautomates/claude-video
• Start with transcript for videos with subtitles or efficient for a fast visual pass; save token-burner for cases that justify the cost of many frames. — https://github.com/bradautomates/claude-video
• Keep frame deduplication enabled: by default it discards nearly identical frames before applying the limit. — https://github.com/bradautomates/claude-video
• Do not append comments after the WATCH_DETAIL value: this previously caused a silent fallback to the default mode, fixed in commit 83da59f. — https://github.com/bradautomates/claude-video/commit/83da59f

Superseded by this¶

• 2026-04-24 — the standalone commands/watch.md wrapper and non-self-contained skill layout were replaced by the skills/watch/ package in version 0.2.0.
• 2026-04-24 — the Windows instruction using python3 is obsolete: on Windows, run scripts with python.
• 2026-04-24 — the early fixed limit of --max-frames=80 was replaced by limits tied to the detail mode.

Still unknown¶

• The dated event provides only the repository URL. No first-party release or commit dated 2026-07-27 was found, so there is no source-backed addition for that date.
• The schema lacks event_findings and new_events fields; the 2026-06-29 release stays in what_changed rather than as a separate structured event.

Sources¶

Agent brief¶

• Subject: project:claude-video, thread claude-video-development, 1 dated events 2026-07-27 → 2026-07-27.
• Practical note: From 2026-07-27, treat the linked bradautomates/claude-video GitHub repository as the source to inspect for implementation and usage details.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.