What does professional video editing actually look like, step by step? This preview dataset holds 234 annotated steps across 4 computer-use trajectories, recorded as working editors built vertical short-form social reels in Adobe Premiere Pro.
We gave editors from the Contra network a client brief and recorded them carrying it through to a finished 9:16 reel: importing footage and screen recordings, laying out selects on the timeline, trimming and reordering clips, reframing for vertical, grading with Lumetri, adding captions and transitions, balancing audio, and exporting. The editors narrated their reasoning aloud as they worked, so every step's thought is transcribed from the expert's own voice.
Each step pairs a screenshot with a first-person thought, a structured action, and executable grounding: an Adobe Premiere MCP tool call, a keyboard shortcut, a menu path, or a coordinate click, with a preferred execution path marking the most deterministic option. The format follows the AgentNet trajectory schema, extended with a Premiere action taxonomy and multi-path execution. This structure makes the data directly usable in several training contexts:
- Computer-use agent SFT: long-horizon timeline-editing trajectories with screenshots, actions, and coordinates in a format interoperable with existing GUI-agent tooling.
- Reasoning mid-training: pair screenshots with narration-grounded chain-of-thought to teach models how experts actually plan, act, and correct course mid-edit.
- Tool-use and function calling: every step carries a structured tool_call, so agents can learn to reach the same goal through an API instead of pixel clicks.
- Benchmarking and evals: a human expert baseline for long-horizon creative workflows, executable end-to-end through the Premiere MCP.
The dataset was recorded in July 2026 on macOS. It is released under the CC BY 4.0 license on Hugging Face.