Inspiration
I shoot drone footage, but when I review it I do not naturally think in editing commands, keyframes or colour-node values. I think in feelings: “this should feel like one continuous flight,” “let this valley breathe,” “the snowy peaks are getting lost in the clouds,” or “this fast rotation feels like an eagle rolling its eyes.”
Turning those observations into precise changes inside DaVinci Resolve was slow. I wanted to see whether Codex could become the bridge between my intuition and the technical edit—not by replacing my taste, but by understanding the footage, translating my words into bounded, verifiable operations, and letting me remain the director.
That became Edit by Feel.
What it does
Edit by Feel connects Codex running GPT-5.6 Sol with DaVinci Resolve Studio 21. I can describe a problem in plain English, and Codex can inspect the real project, identify the exact timeline and clips, analyse the footage, perform a bounded edit, and verify what changed.
The workflow can study DJI telemetry such as GPS, altitude, speed and direction; measure visual properties such as sharpness, exposure, colour and motion; compare shots over time; and inspect whether two cuts will flow naturally. It then uses Resolve’s native scripting interface for precise operations with stable project, timeline and item IDs, with explicit track coordinates.
Resolve’s scripting API does not expose everything. For those genuine gaps, Edit by Feel uses guarded computer control: observe the Resolve window, make one controlled gesture, observe again, and verify whatever can be checked through native scripting. It does not pretend that every creative control is available through an API.
Every substantial experiment follows a safety loop: save, duplicate or create a local version, change one thing, watch it, then approve or roll back.
How I built it
Before Build Week, I had an experimental bridge with only two tools: a connection check and unrestricted Python execution against Resolve. It proved that external control was possible, but it was too broad and unsafe to be a real editing product.
During Build Week, I rebuilt that idea as a Codex plugin with 20 named, bounded tools for project inspection, timelines, clip placement, track state, colour, audio, Fusion, markers, deletion, export and rendering. Inputs use closed schemas, destructive operations are isolated, and timeline mutations can be protected by snapshot hashes. The current safety and protocol suite has 44 passing tests.
I also developed the footage-intelligence workflow around the plugin. It parses DJI telemetry, calculates visual and motion measurements, uses optical flow to distinguish camera movement from high-frequency shake, builds temporal views and seam comparisons, and helps rank possible shots. The measurements narrow the search; they do not make the final aesthetic decision.
For elements that are more reliable as prepared assets, such as frame-exact title animations, I used transparent ProRes 4444 overlays and Fusion. For delivery, the workflow records timeline identity, render properties, hashes and independent quality checks.
The Phyang proof
I tested Edit by Feel on my own drone footage from Phyang in Ladakh. The source set contained seven files, and the finished film became a 64.064-second journey from a human rooftop opening into the valley, layered ridges and a snowy high-altitude finish.
The first versions were technically assembled but felt like a collage. That is where the collaboration became useful. I could say that a circular shot was lasting too long and straining my eyes, that a blink transition was too fast, that one cut had a different sky and mountain shade, or that the continuous drone-air sound felt like a distracting “hush.” Each observation became an isolated experiment.
Some automated transition and retiming attempts made the film worse, so we rolled them back instead of defending them. Stable “eagle-view” shots were allowed to breathe. Colour was matched shot by shot, with a time-ramped correction where the composition changed inside one clip. Snowy peaks received local separation from the surrounding clouds. The distracting air layer was disabled while preserving it for reversal. The titles were refined into a restrained “PHYANG” opening and “THANK YOU / FOR WATCHING” close.
The final master was rendered from DaVinci Resolve Studio as ProRes 422 HQ. Delivery QC decoded all 1,920 frames and found no problematic black or frozen sections.
Challenges
The hardest challenge was not calling Resolve. It was respecting the difference between what can be measured and what must be felt.
A shot can score well for stability and sharpness but still be exhausting because it rotates for too long. Two frames can appear similar while the speed and direction across their cut feel wrong. A mathematically smooth transition can still place too much pressure on the viewer’s eyes. Colour matching also cannot be reduced to copying one grade across every shot, because sky, vegetation, rock, snow and changing light need different treatment.
The other challenge was reliability. Resolve’s native scripting API is powerful but incomplete, while unrestricted screenshot clicking is fragile. The hybrid approach—native scripting first, guarded UI only for gaps, followed by verification—was the practical answer.
What I learned
My biggest learning is simple: the math nominates; playback and human taste decide.
I also learned that “cinematic” does not mean adding more effects. It usually means better continuity, restraint and emotional control. Shot duration should follow visual richness and viewer comfort, not a uniform trailer rhythm. Transitions should normally disappear into the movement.
Most importantly, AI works better here as a reasoning engine and precise pair of hands than as an automatic director.
What’s next
Next I want to package a small judge-friendly sample workflow, expand the bounded support for retiming and reusable transitions, create a reusable colour-look library, and improve how editorial preferences are carried from one project to the next.
The long-term goal is not a one-click video generator. It is an editing collaborator that learns how a person tells stories, works directly inside professional software, explains its choices, and keeps every creative decision inspectable and under human control.
Log in or sign up for Devpost to join the conversation.