Feed it a replay and its telemetry. It finds the moment, cuts it vertical, re-shoots it, and lays the numbers over the top.
driftclip is a tool built with and for Ian to automate short-form drift clips from Assetto Corsa. A telemetry sidecar captures what the car is doing; a moment detector reads that trace and decides where the clip is.
The mechanism is a chain of four phases, each feeding the next: telemetry in, a detected moment, an ffmpeg cut into a vertical frame, a CSP scripted-camera pass that re-shoots the replay, and finally the telemetry overlay and branding on top.
The hard part isn't the cutting. It's the detector. A drift has a peak and the clip has to land on it — but what counts as a good clip is Ian's call, not the code's. So the detector gets tuned only after he has said what he's looking for.
The overlay doesn't get its own data model. It shares the telemetry schema with YAW, the drift telemetry project for Ian's competitive drift car, so one definition of the data serves both the clips and the instruments. Don't invent a second one.
It lives in a Python repo and is built in Claude Code. Started September 2026.
Detect, cut, re-shoot, overlay. Three are built; the re-shoot is the one still to come.
Sidecar + detector
A telemetry sidecar captures the run alongside the replay. A moment detector reads the trace and picks the moment worth clipping. Built.
Vertical cuts
ffmpeg cuts the moment out of the replay into a vertical, short-form frame. Built.
Scripted-camera re-shoot
A CSP scripted camera replays the same moment and re-shoots it. Next.
Overlay + branding
An angle gauge with its held peak, speed and gear, burned into every clip on the schema YAW already defines. Built; branding is still to come.
One Windows app, three of four phases built. It is still waiting for a real session in the sim.
Exists. One download for sim drivers: start it, drive, and each session is recorded through OBS, scored, and cut into 9:16 clips with the telemetry overlay on, to keep or bin from its own page. Tested end to end in a demo mode that stands in for the sim and OBS.
And a way to find out whether the detector is any good. Reviewing the clips already means deleting the ones that miss, so whatever is left in the folder is a verdict — with the detector’s own ranking in the filenames. A comparison step reads that and reports how much of what it chose was wanted, and whether the ranking put the good ones near the top. Where it missed something, it traces the telemetry in that window and names the threshold that blocked it, so tuning is a number to change rather than a guess.
Not yet. A run on Windows next to Assetto Corsa and OBS with a real session. Then phase 3, the CSP scripted-camera re-shoot, and the branding on top of the overlay.
What a "good clip" looks like, in Ian's words — asked before any detector tuning.
Three programs: one listens, one judges, one cuts.
- 01 · LOGGER
Listen to the sim
Reads Assetto Corsa’s shared memory sixty times a second, and starts and stops the OBS recording itself so the telemetry and the video share a clock.
- 02 · SCORER
Find the moments
Turns a session into ranked, typed moments with video in and out points. Four kinds: sustained, transition, save and spin.
- 03 · CUT
Cut them vertical
The top fifteen become 9:16 clips, with a contact sheet to choose from. Keep the ones that land, delete the rest.
Every threshold lives in one config, and a test builds a synthetic session — grip, sustained, transition, save, spin — and checks all four kinds are found, so tuning doesn’t need the sim running. The row format is YAW’s too, deliberately: what reads a sim session reads a real one. Still open, and written down in the repo: tandem proximity, since shared memory only exposes the player’s car.
Here is that synthetic session. Move the threshold and watch moments appear and vanish.
Slip angle over thirty-two seconds — grip driving, a long slide, a transition through
zero, a save, and a spin. The orange bands are what the detector would cut. The dashed line is
sustained_angle, the one number that decides how much of a slide counts.
The session is synthetic — it is the repo’s own test fixture, not a real lap. The detection is not: each threshold setting was scored by the real detector ahead of time, and the browser only draws the answer. A second copy of the thresholds living in JavaScript would drift from the one that does the work.
- The drift-clip project page now runs the detector in the browser: a synthetic session, the moments it found, and a threshold you can move to watch them appear and vanish.
- The drift-clip detector can now be scored against what a human actually keeps — and where it misses something, it names the threshold that blocked it.