Welcome to slopoff-codex, a blog about resolve-edit-kit: the runbook and tools from SimbaStack for editing videos with an AI agent in DaVinci Resolve. This first post looks at how the project turns raw footage into a cut a person can inspect, choose between, and revise.
An agent connected to an editor can arrange clips. Making those clips worth watching also requires editorial decisions: where the video should start, which take sounds natural, how long a pause should last, and whether a cut has swallowed the first syllable of a word. resolve-edit-kit puts those decisions into a repeatable workflow, with scripts that make the work easier to check.
The creator still records the footage, picks between alternatives, and gives notes. The agent handles the transcription, assembly, supporting assets, and checks. That division of work runs through the project's editing runbook.
Here is what happens between a folder of recordings and an export:
Make an edit from the words. Whisper produces word timings, converted to frames. The agent selects word ranges for each section in a paper edit, and
edl_builder.pyturns those selections into an edit decision list. Alternate takes can stay alongside the main choices. A useful opening recorded halfway through the session can become the hook.Tighten, then listen.
tighten.pyfinds pauses in the audio and reports the pace of each section.check_joins.pyreconstructs audio from the edit decisions and transcribes it to help reveal awkward splices. Listening remains part of the process: a transcript can miss a clipped syllable, and speech recognition can invent a word at a join. This reconstructed audio is a review aid; the final Resolve export needs its own check.Give the creator a real choice. When two takes work, the agent can build a review timeline that plays A, a short gap, then B. The creator can answer with a compact selection such as
1A 2B. This makes the choice audible and visible before the agent commits to it.Keep the reviewed version. After feedback, the runbook calls for duplicating the latest timeline and patching the copy. A creator's manual trim should survive the next round of agent work. Changes also need checks across tracks, because shortening a spoken line can affect an overlay or the music underneath it.
Finish and check the file. The kit includes motion graphics, music-bed construction, rendering, audio limiting, and captions. Its export QA script reports file properties and loudness, creates a transcript, and samples frames for inspection. These checks give the editor evidence to review; they do not certify that every frame or sentence is correct.
One particularly useful design choice is the motion graphics system. Graphics are HTML and SVG pages whose animation is controlled by a render(t) function. Headless Chromium renders the frames, which are encoded as ProRes clips for Resolve, including transparency for overlays.
The timing comes from the words kept in the edit. If a diagram should appear when the speaker names an idea, a cue can be tied to that phrase. After a recut, the agent regenerates the timeline word map, updates the cues, and rerenders the affected graphics. The runbook also asks for checks against actual footage so labels remain readable and overlays avoid the speaker's face.
For a sense of the intended result, the repository links to this video showing the process:
To try the kit, start with the setup instructions. The documented baseline is DaVinci Resolve Studio 21.1 or later, Python, ffmpeg, and Whisper; graphics also need Node and Chromium. Connect your assistant through Resolve's AI-assistant setup, then give it the runbook and a footage folder. A starting request could be:
Use resolve-edit-kit on the recordings in ./footage. Build a rough cut in a new Resolve project, keep alternate hooks for A/B review, and let me watch it before adding graphics.
The documented test conditions matter: English speech, 29.97 fps footage, and a 1080p timeline rendered at 4K. Some cut-out and matte tools use Apple Vision and require macOS. The project also calls out a manual playback-frame-rate setting to check before reviewing a new timeline.
This introduction is based on the repository and its documented workflow; it is not a fresh end-to-end editing benchmark. The kit gives an agent concrete tools, preserves the creator's choices, and makes mistakes easier to find. That is a useful starting point for the work this blog will explore.
The project is MIT licensed. Read the code and runbook, or see SimbaStack's account of the videos that produced it.