Step 1: record once in OBS
You record a single widescreen session with OBS Studio, teaching or demonstrating as normal. That one file is the source for everything else. Flubs are fine: if you say something wrong, say that you will do it again and repeat the sentence. The exact recording settings are on the OBS settings page.
A small helper also logs where your mouse is and when you click, while you record. It never records keystrokes. That log is used later so the shorts can follow your cursor and so a silent demonstration is not mistaken for dead air.
Step 2: AI cuts the long video, and you approve the report
A local speech-to-text model in the Whisper family transcribes the recording with word timings. From that transcript and the audio, Claude proposes cuts: long silences, filler sounds such as "um", and retakes where you repeated a sentence.
Checkpoint 1. Before anything is rendered you get a cut report: what was removed and why. You say yes, or veto any single line. Only then is the video rendered.
The finished layout is a standard one: a cold open (a short clip of the video itself, picked as the hook), a brief logo sting, the content, and a 20-second end-screen plate with empty zones where you place YouTube's end-screen elements in YouTube Studio. The voice gets a light clean-up, and music sits under it at a lower level.
Step 3: titles, description, chapters and captions
With the final edit done, Claude writes the title options, the description, the chapter list and a captions file. Chapter times are calculated against the final video, after the intro, so they land in the right place. If a clip shows simulated trading results, the on-screen label and the required disclosure text are added from a reference file, which is a good example of a rule that is easy to forget and easy to automate.
Step 4: AI picks the shorts, and you approve the pans
Claude chooses four to six moments that make sense on their own and turns each into a vertical 9:16 clip, panning across the widescreen frame to follow the action. Checkpoint 2: you watch a preview of every pan, with a box showing what each short will display, and you say yes or correct it. Details are on the shorts page.
Step 5: review and tweak in Premiere Pro
Every long-form edit is also written as an editable Premiere Pro timeline, with the screen, voice, desktop sound, music and effects on separate tracks and a marker on every edit with its reason. Any cut can be undone by dragging a clip edge, because the whole recording is still on the timeline. See human in the loop.
Step 6: you post
The finished video, the shorts, covers and post captions land in a folder. A person uploads them. The pipeline does not post, and that is a deliberate rule, not a missing feature.
Why there are two checkpoints
Cutting and panning are the two places where the AI makes judgment calls that change what viewers see. A wrong cut can remove the one useful sentence; a wrong pan can show the empty corner of the screen. Reviewing a short report and a set of previews takes far less time than doing the edit, and it keeps the final responsibility with you.
What is tested, and what is not
The pipeline has been run end to end on synthetic test recordings, with audio and video sync measured within about a millisecond. A real recording, the Premiere Pro import and the NVIDIA GPU paths have not been tested yet. This site will be updated when they are.
Frequently asked questions
Does the AI post my videos for me?
No. The pipeline makes the files and a human posts them. Nothing is uploaded automatically, and any future upload tool would need an explicit yes for each video.
Do I need to record separately for the long video and the shorts?
No. You record one widescreen session in OBS. The long YouTube video and the vertical shorts are both made from that same recording.
What happens if I make a mistake while recording?
Keep going. Say you will redo the sentence and say it again. The cutter keeps the clean take, and the cut report shows you what it removed so you can veto any line.
Is the whole workflow proven?
Not yet. It has been tested end to end on synthetic recordings. A real recording, the Premiere Pro import and the NVIDIA GPU encoding paths are still untested.