Produce

Autonomous Rough Cut Engine

Usage

/produce:produce /path/to/raw-video.mp4

What it does

Takes a raw face cam recording (any length, with retakes) and outputs a clean rough cut in original resolution. Removes:

• Retakes (keeps only the last take of each section)
• False starts and fragments
• Side conversations (talking to the AI about the script, not recording)
• Internal repetitions (e.g., double CTA starts)

Adds:

• 30ms audio fades at cut points (no clicks/pops)
• Room tone bed at 12% volume (fills micro-gaps naturally)
• Smart gap handling (preserves natural breaths between sentences)

Requirements

• Claude Code
• Python 3.10+
• uv (Python package manager)
• ffmpeg
• An OPENAI_API_KEY environment variable (used for per-chunk Whisper transcription)

How to run

1. Get the video path from the user argument
2. Verify file exists: check with ls
3. Run the pipeline from the bundled engine:

cd /engine
OPENAI_API_KEY=sk-... uv run python roughcut.py ""

1. The script will:

• Extract audio and analyze waveform (0.3s threshold)
• Transcribe each speech chunk individually via OpenAI Whisper API
• Find the last complete run-through
• Detect and remove retakes (rapidfuzz + key word overlap + partial containment)
• Build the rough cut with audio fades and room tone
• Cross-verify by re-transcribing the output
• Open the final file automatically

1. Report results to the user: duration, chunks kept/skipped, any issues detected

Output

• File saved to: exports/produce/_/_roughcut.mp4
• Original resolution (1920x1080 or whatever the source is)
• CRF 16, AAC 256kbps (near-lossless quality)
• Ready to import into Descript, Premiere, or any editor for final assembly

Architecture

• Waveform-first: 10ms RMS energy windows find exact speech boundaries
• Per-chunk transcription: Each speech chunk sent to OpenAI Whisper API separately (prevents merging retakes)
• Last-take rule: You record multiple takes, keep repeating until satisfied, then move on. Last take before a topic change = keeper.
• Three-layer retake detection:

1. Text similarity (rapidfuzz tokensortratio > 60)
2. Key word overlap coefficient > 0.60
3. Partial containment (short chunk's first words appear in a later chunk)

• Smart gap handling: Natural breaths between kept chunks preserved, gaps where content was removed get room tone fill

Settings

These defaults are tuned for a natural conversational speaking style. Adjust in engine/roughcut.py if your cadence differs:

• Silence merge threshold: 0.3s
• Min chunk duration: 1.0s
• Fragment max duration for partial containment: 3.0s
• Breath gap max: 1.5s
• Room tone volume: 12%

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: manthanpatelll
• Source: manthanpatelll/leadgenman-video-skills
• License: MIT

Install and usage instructions live in the source repository linked above.