Aaron Helwig

Cutwright AI video editing

I built Cutwright to give AI agents a usable editing workflow for client video work. It connects source inspection, transcripts, editorial decisions, motion graphics and render review, so an agent can work through an edit without losing track of which footage and words its decisions refer to.

Keeping the edit attached to the words

A source transcript and a finished edit do not share the same clock. Remove a sentence near the beginning and every later moment moves. A graphic planned against the original timestamp can land on the wrong phrase even when both the graphic and the cut look individually reasonable.

I built a mapping from source words into the cut's output timeline, preserving the identity and timing of each selected word. Visual events can refer to those anchors, and validation checks whether an anchor still belongs to the chosen source and selected passage.

Separate the decision from the render

The system separates reusable client style and production instructions from each video's source media, cut, visual plan, rendered graphics and review records. A visual idea can be revised without confusing it with the file currently placed in the timeline. The agent can inspect source footage, an individual graphic or the final composite as different surfaces.

Pipeline

Transcribe -> Pack -> LLM Reasons -> EDL -> Pre-render Lint
                                           |
                                           v
                         Optional Visual Slots -> Render
                                                   |
                                                   v
                                      Render QC -> Visual Self-Eval
                                                   |
                                                   v
                                         issue? fix + re-render

The frame between two shots

One particularly useful check came from a small but visible failure: a transition could uncover the frame at a camera cut and produce a flash. I added boundary checks around opaque transitions so the edit could identify that timing error before treating the render as ready.

The workflow produced a full-length rendered edit. Alongside the output, it generates review material and records what still needs attention, so the next editorial pass can start with specific moments to inspect.

Cutwright began from Browser Use's video-use project. I retained useful rendering and transcription foundations while developing the project structure, word mapping, visual-event validation and review machinery around the needs of this editing workflow.