Talking Reels: 100% Edited by Claude AI

Edit take talking-head saya jadi Reels gaya "greget": potong bagian saya diam, kasih zoom, teks, caption kata per kata, motion graphics, SFX dan musik. Semua diedit lewat kode, tanpa CapCut atau Premiere. Tutup dengan ajakan follow @yogi.ef.

Starting point

One raw take: me telling the camera that the video you're watching was edited by AI, not CapCut or Premiere. The agent did the rest.

How it works

• Word-level transcription (faster-whisper). Silences cut out, leaving 16 jump-cut segments.
• A per-frame face track lets the framing change on every cut and every ~2 s (zoom 1.0 to 1.36).
• A person matte (MediaPipe) puts giant words like EDITAN, CLAUDE and OKE behind the body.
• 13 overlays in the top zone and 6 full-screen inserts with a face PiP: steps, Claude, code, render, montage.
• Word-by-word captions. The spoken word sits on a blue pill, and keywords are set in gold serif.

v1 (2026-10-28)

The problem: the full 66 s audio played under a video already cut down to 58 s, so the voice drifted further at every cut, up to 4.9 s. It also had only 8 SFX and no music.

v2 (2026-10-29)

A revision after the feedback "audio out of sync, not punchy, too few SFX, no music":

• The voice is cut per EDL segment, and cross-correlation shows 0 ms lag on every segment.
• Procedural 140 BPM F-minor trap music, copyright-free. Its drops land on "100% AI", "CLAUDE" and the montage, and it is ducked −7 dB under the voice.
• SFX raised to 238 cues, and the camera pulses on every kick.
• Mastered to −14.1 LUFS.

Cover

Three rounds. v1 used the face. v2 was too close to the reference. v3 is a retro TV whose screen shows real edited frames, in the video's own palette and fonts.