When a test run calls for 20 to 30 clips, manual editing becomes the bottleneck in the entire hypothesis funnel. Cuts, titles, resizing for placements, endless tweaks — a full day can disappear even after the hook and concept are ready. To see how much of that routine can actually shift to AI, two Claude-based setups received identical source files and five tasks from a real creative production cycle.
Two setups, not one
The authors skipped ready-made paid 'editor-in-a-box' agents. Instead they assembled two nominally free stacks where Claude acts as the brain and separate tools handle the actual video work. Nominally free because the core software costs nothing, but a Claude subscription still has to be paid. Several components carry licensing caveats.
First stack: Claude + OpenChatCut. OpenChatCut is an open video editor with a timeline and dialog agents in one window, unrelated to the commercial ChatCut. Claude connects via an external MCP server: the editor launches it on a local address, then a terminal command runs claude mcp add –transport http openchatcut http://localhost:5199/api/external-mcp/mcp, a skill installs with npx skills add 0xsline/OpenChatCut, and the skill router loads 26 specialized skills. If the editor opens in a browser, the source folder must be declared in AGENT_IMPORT_ROOTS in the .env.local file. In manual mode Claude drafts edits and the user confirms them in the editor. In auto mode edits apply at once, but generation, export and project deletion cannot run through external MCP — the final MP4 exports by hand. Constraints include early unsigned macOS builds, an MCP bridge built for one user on one machine, and no Claude subscription login — API key only. Building from source requires Node.js 24 and npm.
Second stack: Claude + FFmpeg + Remotion + Whisper. No timeline here. Claude writes and runs commands and code, each tool closes its own segment. FFmpeg handles file operations: transcoding, trimming, joining, scaling, filters; ffprobe shows file parameters. OpenAI's Whisper, trained on 680,000 hours of multilingual audio, delivers text and timecodes; it works with 99 languages, outputs timestamps at word or segment level, models range from tiny to large. Remotion assembles the design: video is described in React code, covering titles, overlays, animation, transitions, data-driven templates. Free for individuals and companies up to 3 people, paid terms beyond that. The pipeline: Whisper transcribes the source, Claude reads the transcript and decides which fragments to take, FFmpeg cuts and joins, Remotion applies titles and graphics.
What five tasks revealed
Test source files came from real work: a long shoot recording, a finished working creative as reference, and materials for substitution. The target output was video creatives ready for moderation without rework. Every manual correction was logged separately.
Cutting raw footage. The agents received a long stream recording of slot gameplay and were asked to assemble three 9:16 ad videos for Facebook and Instagram with different hooks and CTAs. The first stack (OpenChatCut) delivered almost without notes in roughly 15 minutes: the agent cut out gameplay and streamer reaction areas, placed fragments on a bright background, no manual fixes required. The second stack (FFmpeg + Remotion + Whisper) spent 11 minutes but burned nearly twice the tokens and in the first iteration cropped the horizontal video incorrectly — it crudely grabbed the center portion. After a series of clarifying prompts the agent inserted the source video into the top of the vertical creative and filled the free space with a blurred image.
Editing to a reference. The agents were given a finished creative and asked to replicate structure, pacing, design and graphic elements on different material. Both pipelines handled the first iterations identically: they simply tore fragments from the reference and the new creative and merged them — the first overlaid cut-out areas on top, the second split the screen into 4 parts. After clarifying prompts the first stack's quality stayed low, the second's agent gradually drifted from the task, hallucinated, and the videos got worse.
Adapting a creative for another GEO. A Lucky Jet creative for an Uzbek audience served as the base; an Aviator clip had to be adapted for Polish with titles translated into Polish. The first agent (OpenChatCut) gave a noticeably better result: it increased the speed of the multiplier growth segment by roughly 5 to 6% so the withdrawal moment and the press of the 'ZAKŁAD' button occurred around the 9.5-second mark, matching the reference. It added a warp-effect tunnel with RGB channel splitting and blur, while the streamers' webcam window stayed in place. In the final frame the translated heading appeared line by line in the yellow-white font styling of the original creative.
Overall score and limits
The Claude + OpenChatCut setup performed better on tasks requiring precise timing and design element work. The second stack spent more tokens, made more first-iteration mistakes, and held the task less well during long dialogs. Editing to a reference proved the hardest task for both agents: both leaned toward mechanical copying of elements rather than reproducing the style on new material. Tasks such as cutting source footage by semantic fragments were closed by the first stack practically without human input.
