recut-reference

Import the video with recut.media.import, mark it with the reference role attribute, then read it using probe, contactSheet, and frames. The understanding is written back to the asset's content field and ref.* attributes such as ref.hook, ref.beats, and ref.transferable.

Use yt-dlp with browser cookies to probe and download the video to your workspace, then import the file with recut.media.import. Platform pages are not downloaded directly; only files imported into the media library count as completed.

Yes. Understanding a reference video requires a model that can read images. If the current model lacks vision, the Skill stops immediately and asks you to switch models rather than substituting transcripts or cut points for actual frame reading.

It writes controlled ref.* attributes: ref.summary, ref.format, ref.hook, ref.beats, ref.systems, and ref.transferable, plus role and url for provenance. Long-form understanding with timestamped details goes into the asset's content field with inline media references.

Zero frames usually means the requested interval exceeds the video duration, so run probe first to get the real length. Other causes include missing ffmpeg or PySceneDetect dependencies, which must be prepared through understand.prepare rather than manual installation.

The Skill only reads and records understanding; it does not decide what to replace in a remake, which belongs to the director skill, nor does it execute the clone. Analysis artifacts stay workspace-level and are never attached to projects.