Claude Code Automates Full Video Editing Pipeline
Build a folder-based system in Claude Code using Whisper and FFmpeg: auto-transcribe raw videos, cut mistakes/silences, add text hooks/captions, output ready shorts—frees 15-20 hours/week for more content creation.
Pipeline Setup: Local Tools for End-to-End Editing
Create a folder with 'raw' and 'outputs' subfolders. Use Claude Code in terminal (install via its quick-start paste; drag folder path in). Switch to planning mode (Shift+Tab) and prompt: transcribe with free faster-Whisper (no API key), detect repetitions/bloopers/errors/silences via transcript analysis, cut with FFmpeg, add hook overlay in first 6 seconds (big bold font, top third), burn in captions, export to outputs with date/name. Auto-accept edits to build scripts. Run full pipeline by dropping raw video into 'raw'—processes in minutes, handling errors like apostrophes automatically.
This local setup (Claude API for intelligence, Whisper/FFmpeg for heavy lifting) turns unedited footage into tight shorts optimized for Reels/Shorts/TikTok, trimming silences for high retention and avoiding platform-native captions.
Hook and Caption Techniques: Psychology-Backed Overlays
Text hooks run parallel to spoken hook but differ: intrigue via paradox (e.g., "She was right" or "It cost me everything"), social proof gap ("78,000 people knew this before me"), or confession ("I almost didn't post this"). Position big/bold in top third, first 6 seconds, plain white on solid black background—no opacity.
Captions mimic top creators like Mino Wee (530k IG followers): small/nonchalant font (e.g., Inter), white with black drop shadow for contrast/legibility, break into 2 words early (first 10s for speed), 4-5 words mid-video, center-aligned, line breaks every 15-20 words later. Auto-correct spelling/grammar (e.g., "Claude" not "Claw"). AB test styles—nonchalant boosts authenticity, reduces clutter for better retention.
Integrate by feeding these rules/examples (e.g., Mino's transcript) into Claude prompts during build.
Refinement, Testing, and Daily Scheduling
Test end-to-end first: drop video, run pipeline, review output (e.g., fix cropping, jumping captions, compression via FFmpeg flags for quality). Iterate conversationally—Claude self-fixes (e.g., re-transcribe post-edit for timing). Read its logs/responses to learn error patterns, speeding future builds (e.g., preempt apostrophes).
Schedule in Claude desktop app: open folder, prompt routine for 9AM daily—scan 'raw', process in parallel (CPU-heavy, overnight ideal) or sequential, output to 'outputs', move raw to 'processed'. Computer must stay on/awake. Extend with tools like Blowtato for auto-publishing.
Outcome: Scales content (Duncan grew 110k followers, 6-figure agency in 12 months; 2k community members automate in <3h/week). Trade-off: Local processing ties to your machine; read Claude outputs to partner effectively without coding knowledge.