How do you make a rough cut from a script with Claude Code?
Give the agent the script, one visual line per shot, and let it call a local footage index for each line. It comes back with a best clip and two alternatives per line, saves the result as a project you can trim by hand, and exports an MP4 or an FCPXML that opens in Resolve, Premiere or Final Cut with your original files linked.
Published October 15, 2026 · Field notes
A rough cut is mostly retrieval. For each beat of the script you go looking for a shot that fits, drop it on the timeline, and move on. The judgement is in the last ten percent; the first ninety is searching. That makes it a good job for an agent, as long as the agent has something faster than scrubbing to search with.
The setup
This example uses Claude Code and Reelary, which is our app. Reelary keeps a local index of your footage and exposes it to agents over MCP. One command installs the app, connects it to Claude Code and adds the skills:
$ npx reelary Then, in Claude Code: "Index ~/Movies/Lisbon with Reelary." Indexing runs in the background on your machine; for a few hundred clips that is minutes, for a whole archive drive leave it overnight. You only do it once per folder.
Step 1: write the script one shot per line
Search works best on visual lines. "We arrived on Tuesday" describes nothing a camera saw; "Tram climbing a narrow street" does. If you paste prose, the rough-cut skill asks the agent to rewrite it into short visual beats and show them to you first. A 30-second travel opener might be:
Sunrise over the rooftops. Yellow tram climbing a narrow street. Coffee being poured at a café counter. Waves hitting the rocks below a lighthouse. Crowd on a viewpoint at sunset.
Step 2: ask for the cut
"Rough cut this script from my footage, 9:16." The agent calls reelary_rough_cut with the five lines.
For each line Reelary searches the index and returns the best clip and two alternatives, each with exact start and
end times, capped at six seconds by default. The agent reads the draft back to you line by line, with the
description of each pick, and flags any line that found nothing.
This is where the agent earns its place. You can say "the tram shot is too dark, use the second one" or "make the whole thing punchier" and it swaps alternatives or shortens shots without you opening anything.
Step 3: save it and look
The agent saves the draft as a Reelary project and can open it in the editor. Here you take over: play it through, nudge a trim, swap a shot from the clip bin, add music and captions. The agent is good at finding candidates; it has not watched the shots the way you will.
Step 4: export
Two ways out, both available to the agent:
- An MP4 with
reelary_export_video, rendered on your machine. Good for a social cut or a client preview. - A timeline with
reelary_export_nle: FCPXML for Final Cut and Resolve, XML for Premiere, EDL as a fallback (Pro). The project points at your original camera files, so grading and sound start from full-quality media, not a render.
"Export it for DaVinci Resolve to ~/Desktop/lisbon.fcpxml" is the whole instruction. Import it in Resolve and the five shots are on the timeline in order, linked to the files on your drive.
Where it stops
- Search finds content, not quality. It will find a tram; it will not know which take is in focus.
- No beat matching or story sense. Order comes from your script. Pacing beyond shot length is your call.
- Simple edits only in Reelary. Grading, effects and audio mixing belong in the NLE you export to.
- The agent sees text. Paths, times and descriptions reach its model provider; the video does not. Details in what an AI agent sees of your footage.
When an agent is overkill.
Reelary's own editor does the same thing without a chat: type the script, get a shot per line, trim, export. The agent version is worth it when the cut is one step in a bigger task - writing the script, making three aspect ratios, filing the export - or when you already live in Claude Code all day.
Questions people ask about this
Does the agent edit the video itself?
No. The agent chooses and orders clips by calling tools; the cutting, rendering and FCPXML writing happen in the Reelary app on your machine. The agent never handles the video data.
Can I use Cursor or Codex instead of Claude Code?
Yes. The same MCP server and skills work in Claude Desktop, Cursor, Codex and other MCP clients. The prompts in this article work unchanged.
Do I need Pro for this?
Not for the rough cut or an MP4. Free exports are 720p, watermarked, up to 60 seconds and five a month. The FCPXML, XML and EDL export for Resolve, Premiere and Final Cut is part of Pro.
What if a line has no matching footage?
The draft lists it as an empty scene. The agent suggests a different wording, which often helps because search works by meaning; if you never shot it, no wording will find it.