Testimonial Edit

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)