testimonial-edit

name: testimonial-edit version: 1.0.0 description: "Edit testimonial videos into polished social clips. Triggers on: 'testimonial edit', 'edit testimonial', 'testimonial video', 'cut testimonial', 'testimonial clip', 'social proof video', 'customer video edit', 'case study video'."

Testimonial Video Edit

Transform raw customer testimonial videos into polished, short social clips with AI-driven content selection, speaker intro cards, and multi-format output.

When to Use

• User has a raw testimonial video (customer interview, case study, social proof)
• Need to cut down a long testimonial to the best 30-90 seconds
• Need multi-format output (square + vertical, or landscape + square)
• Need speaker name/title/company intro card overlay

Before Starting

Gather from User (ask if not provided)

Prerequisites

• FFmpeg installed (brew install ffmpeg)
• DEEPGRAM_API_KEY in .env.local (transcription)
• ANTHROPIC_API_KEY in .env.local (AI content selection)
• Remotion dependencies installed (npm install)

Pipeline (10 Stages)

[1] PROBE ─→ [2] TRANSCRIBE ─→ [2.5] IDENTITY CHECK ─→ [3] AI SELECT ─→ [4] SILENCE ─→ [5] CROP
                                                              ↓ (review gate)
[6] APPLY EDITS ─→ [6.5] CAPTION REVIEW ─→ [7] ENHANCE ─→ [8] MANIFEST ─→ [9] RENDER ─→ [10] QA

Workflow

Step 1: Gather Input

Ask the user for the video path, speaker info, and preferences. Then run:

tsx scripts/process-testimonial.ts <video-path> \
  --name "Ray Lee" \
  --title "CEO" \
  --company "Zero2.AI" \
  --caption-style pill

This runs stages 1-3 and pauses at the interactive review gate.

Step 2: Review AI Content Selection

The pipeline outputs scored snippets to snippets.json and prints a summary:

  KEEP [9/10] (12.3-18.1s) We saw a 40% increase in qualified leads...
  CUT  [3/10] (0.0-4.2s)   Hi everyone, thanks for having me...

Present the AI's edit plan to the user. They can:

• Approve as-is → re-run with --auto-approve
• Adjust keepSegments in snippets.json → re-run with --auto-approve

Step 3: Render

Once approved, the pipeline continues through stages 4-10 automatically:

tsx scripts/process-testimonial.ts <video-path> \
  --name "Ray Lee" \
  --title "CEO" \
  --company "Zero2.AI" \
  --caption-style pill \
  --auto-approve

Step 4: QA

Pipeline auto-extracts QA frames. Open outputs in Preview:

open output/testimonials/<job-id>/output-square.mp4
open output/testimonials/<job-id>/output-vertical.mp4

Step 5: Save to Nextcloud

Copy finished videos to ~/Nextcloud/Testimonials/ with descriptive titles:

mkdir -p ~/Nextcloud/Testimonials
cp output/testimonials/<job-id>/output-square.mp4 \
  ~/Nextcloud/Testimonials/"<Speaker Name> - <Company> Testimonial (Square).mp4"
cp output/testimonials/<job-id>/output-vertical.mp4 \
  ~/Nextcloud/Testimonials/"<Speaker Name> - <Company> Testimonial (Vertical).mp4"

Naming format: Ray Lian - Sarah2.ai Testimonial (Square).mp4. Nextcloud syncs automatically.

Output Format (Adaptive)

Auto-detected from ffprobe. AI face detection (face-api.js 68-point landmarks) extracts a frame at 2s and computes crownY, chinY, faceCenterX, faceCenterY, leftEarX, rightEarX.

For vertical→square crops: the crop window is positioned so the crown touches the top of the frame. Full source width is preserved (no horizontal cropping, no blur background). This produces a tight, professional framing where the speaker's head fills the top of the square. Override with --crop-y.

AI Content Selection Rules

High-Value (Keep)

• Specific results: numbers, metrics, dollar amounts, timeframes
• Emotional impact: genuine enthusiasm, surprise, relief
• Social proof: client names, revenue figures, team adoption
• Problem/solution: "before" state, comparison to alternatives

Low-Value (Cut First)

1. Introductions and sign-offs
2. Filler and hedging ("um", "you know", "I think maybe")
3. Repeated points (keep stronger version)
4. Tangents and off-topic segments
5. Generic praise without specifics
6. Program pricing / cost details — NEVER reveal what the customer paid to join (e.g. "$1,500", "entry price", "how much it costs"). This is proprietary sales info that should not appear in public testimonial clips.

Sentence Completeness (Critical)

• Every kept segment MUST end on a complete thought
• If a sentence trails off mid-phrase ("I hope you believe...", "and he's really..."), either cut it or ensure it merges with the next segment
• The last kept segment is especially important — the video must end cleanly, not mid-sentence
• Review each kept segment's final words to confirm the thought is finished

Constraints

• Achieve 50%+ reduction from original
• Minimum 20s output (configurable with --min-duration)
• Maximum 90s output (configurable with --max-duration)

TestimonialIntroCard Component

Brand System

Audio Enhancement Pipeline

1. FFT Denoise: afftdn=nr=12:nf=-25:tn=1
2. Highpass 80Hz: Remove room rumble
3. Lowpass 12kHz: Remove hiss
4. EQ 200Hz +1dB: Warmth
5. EQ 3kHz +2.5dB: Clarity/presence
6. EQ 6.5kHz -2dB: De-essing
7. Compressor: Smooth dynamics
8. Loudnorm: -16 LUFS (two-pass with limiter)

CLI Reference

tsx scripts/process-testimonial.ts <video-path> [options]

Options:
  --name "Speaker Name"     Speaker name for intro card (required)
  --title "CEO"             Speaker title
  --company "Acme Inc"      Speaker company
  --caption-style pill      Caption style (pill|pop|karaoke|tiktok|subtitle-bar)
  --no-captions             Skip captions entirely
  --no-music                Skip background music
  --no-silence-removal      Keep all silences
  --auto-approve            Skip interactive review of AI snippets
  --transcript <path>       Use existing transcript (skip Deepgram, saves ~$0.01)
  --crop-y <0-1>            Manual crop Y position (0=top, default: 0.25)
  --min-duration <sec>      Minimum output duration (default: 20)
  --max-duration <sec>      Maximum output duration (default: 90)
  --output <dir>            Custom output directory
  --render-only <manifest>  Skip to render with existing manifest

Output Directory Structure

output/testimonials/<job-id>/
  transcript.json           — Word-level transcript (Deepgram)
  transcript-edited.json    — Re-transcribed after edits
  transcript-reviewed.json  — AI-reviewed captions (corrected words)
  snippets.json             — AI content selection (scores + reasons)
  edited.mp4                — Trimmed video (AI cuts + silence removal)
  enhanced.mp4              — Audio-enhanced version
  cropped-square.mp4        — Cropped for square format
  cropped-vertical.mp4      — Cropped for vertical format (if vertical input)
  manifest-square.json      — Remotion manifest (square)
  manifest-vertical.json    — Remotion manifest (vertical)
  output-square.mp4         — Final rendered square (1080x1080)
  output-vertical.mp4       — Final rendered vertical (1080x1920)
  qa/                       — QA verification frames

Cost Estimate

Autonomy Rules

• Don't ask permission between pipeline stages (except the AI review gate)
• Auto-detect aspect ratio — only ask if ambiguous (very close to 1:1)
• Default to pill captions for testimonials (professional, clean)
• Always pause at AI content selection for review unless --auto-approve
• Skip re-transcription if --transcript provided
• If video is already under 30s, skip AI content selection entirely

Known Gotchas

Files

Reused (not modified)