Skip to content

Latest commit

 

History

History
274 lines (210 loc) · 12.5 KB

File metadata and controls

274 lines (210 loc) · 12.5 KB

Case Nexus — Video Production Guide

Everything you need to produce the 3-minute demo video.


Assets

Asset Path Notes
Title card (16:9, 5504x3072) static/title_card.png Opening frame — "Public defenders carry 500 cases."
Base thumbnail (no text) static/evidence/image-1771034879440.jpeg If you want to re-composite

Software You Need

Tool Purpose Get it
OBS Studio Screen recording (1080p60, separate audio track) obsproject.com
DaVinci Resolve (free) Video editing, text overlays, speed ramping blackmagicdesign.com
Voice Memos (macOS built-in) Voiceover recording Already on your Mac
Alt: CapCut (free) Faster editor if you know it, good text tools capcut.com

OBS Settings

  • Canvas Resolution: 1920x1080
  • Output: Recording → CRF 18, Encoder x264, High profile
  • Audio: Disable desktop audio capture (you'll record voice separately)
  • Source: Window Capture → Chrome window (or Display Capture with everything else hidden)

Browser Setup

  • Chrome, 100% zoom, window sized to exactly 1920x1080
  • Hide bookmarks bar: Cmd+Shift+B
  • Close all other tabs
  • Disable notifications: System Settings → Focus → Do Not Disturb ON

Pre-Recording Checklist

# 30 minutes before recording
rm -f case_nexus.db && python app.py

# Do ONE throwaway Health Check to warm the API connection
# Then restart fresh:
rm -f case_nexus.db && python app.py
  • .env has valid ANTHROPIC_API_KEY with ~$15-20 credits
  • Chrome open, bookmarks hidden, 1920x1080
  • OBS recording, correct window selected
  • Phone on silent, Slack/Discord/Mail quit
  • Do Not Disturb ON

Recording Plan — 5 Separate Takes

Record each section as its own clip. Pick the best take of each. Never try one continuous recording.

Take What to capture Real-time Target after edit
Take 1 Landing page → click Sync Caseload → cases loading ~15s 12s
Take 2 Click a case → click Health Check → wait → results appear 60-90s 50s
Take 3 Click Cascade Intelligence → tool calls stream → brief appears 60-90s 45s
Take 4 Navigate to CR-2025-0012 → click Adversarial Sim → all 3 phases 120-180s 45s
Take 5 Click an evidence image → analysis starts → cut to dashboard ~15s 15s

Record each take 2-3 times. Pick the cleanest one.


The Edit — Shot by Shot

TITLE CARD (0:00–0:03)

Visual: static/title_card.png — fade in over 0.5s, hold 2s, fade to app.

Audio: Silence. Let the image speak.


SYNC + PROBLEM (0:03–0:15)

Time Screen Voiceover Text Overlay
0:03 App landing page appears "They walk into court with minutes to prepare." Fulton County, Georgia bottom-left
0:07 Click "Sync Caseload" "Case Nexus loads every case into Claude's million-token context window."
0:10 Cases populating (speed to 2x) "No RAG, no chunking. The AI sees everything at once." 500 cases • 275,000+ tokens bottom bar

MOMENT 1 — Health Check (0:15–1:05)

Time Screen Voiceover Text Overlay
0:15 Click case CR-2025-0012, detail card shows (silence — let judges read the card for 2s)
0:18 Click "Health Check" "Health Check. All 500 cases in a single Opus 4.6 prompt." 275K tokens loaded near token bar
0:22 Token bar filling, thinking starts Speed to 3x during dead air, 1x when text streams
0:28 Thinking text streaming "Claude uses 60,000 tokens of extended thinking — all streamed live."
0:35 Results appear — 1x speed from here "The AI found Officer Rodriguez in four separate traffic stops — all with contested vehicle searches."
0:45 Mouse slowly over alerts panel "A Fourth Amendment pattern invisible across 500 files."
0:50 Pan to cross-case connections "Speedy trial deadlines, plea disparities, shared witnesses — the full risk profile." Amplify Human Judgment bottom-right
1:00 Pan to risk heatmap, hold 5s (silence — let judges read)

Editing tip: Subtle 105% zoom-in (Ken Burns) on the alerts panel when the Rodriguez finding appears.


MOMENT 2 — Cascade Intelligence (1:05–1:50)

Time Screen Voiceover Text Overlay
1:05 Click Cascade Intelligence (gold button) "Cascade Intelligence. Nine tools. Claude investigates autonomously." 9 tools • 8 rounds • autonomous
1:12 First tool call appears — 1x speed "Claude pulls case CR-2025-0089 for deeper investigation..."
1:18 Second tool call "Now looking up the Georgia statute..."
1:24 Third tool call "Searching for case law precedent..."
1:30 Gap between calls → speed to 2x
1:35 Strategic brief rendering — 1x speed "A genuine agentic loop — Claude decides what matters, gets the information, builds strategy."
1:42 Slow scroll through brief (silence — let judges read key findings)

Editing tip: Add a subtle click/ping sound effect when each tool call badge appears. Gives rhythm. Free at pixabay.com/sound-effects.


MOMENT 3 — Adversarial Simulation (1:50–2:35)

Time Screen Voiceover Text Overlay
1:50 Click CR-2025-0012 → click Adversarial Sim "Three AI sessions chained. Prosecution. Defense. Judge."
1:55 Prosecution phase streaming Speed to 2x during wait, 1x when content appears Phase 1/3: Prosecution top-left
2:05 JUMP CUT → defense phase starts "The defense reads the prosecution's actual arguments and dismantles them." Phase 2/3: Defense top-left
2:12 Split panel visible for 5s (silence — let judges read both sides)
2:17 JUMP CUT → judge panel appears "80,000 tokens of chained reasoning across three perspectives." Phase 3/3: Judicial Synthesis
2:25 Judge output readable for 5s "Prosecution, defense, judicial — each with visible reasoning the attorney can challenge."

On jump cuts: Quick 0.3s fade-to-black between phases. Add a small text card: ~45 seconds of analysis to legitimize the time skip.


EVIDENCE FLASH + CLOSE (2:35–2:55)

Time Screen Voiceover Text Overlay
2:35 Click an evidence item (dashcam image) "Claude also analyzes evidence using vision — dashcam footage, injuries, crime scenes."
2:40 Analysis thinking starts streaming → cut away after 3s
2:43 Dashboard view — slow pan across alerts, connections, heatmap "Every Opus 4.6 capability — million-token context, extended thinking, autonomous tool use, 128K output, vision, chained reasoning — amplifying one attorney's judgment across 500 cases."
2:52 Hold on dashboard, token bar visible Case Nexus centered, then Built with Claude Opus 4.6 below
2:55 Fade to black

All Text Overlays — Quick Reference

Add these in your editor. Use DM Sans or Avenir Next (matches the app).

Time Text Position Style
0:03 Fulton County, Georgia Bottom-left Small, subtle, white
0:10 500 cases • 275,000+ tokens Bottom center bar Semi-transparent bg
0:18 275K tokens loaded Near the token viz bar Small chip
0:50 Amplify Human Judgment Bottom-right Subtle, smaller
1:05 9 tools • 8 rounds • autonomous Bottom center bar Semi-transparent bg
1:55 Phase 1/3: Prosecution Top-left Label with bg
2:05 Phase 2/3: Defense Top-left Label with bg
2:17 Phase 3/3: Judicial Synthesis Top-left Label with bg
2:52 Case Nexus — Built with Claude Opus 4.6 Center Fade in, larger

Voiceover Script — Complete

Record this AFTER editing the screen capture. Watch your edit on loop, then speak to match the timing.

They walk into court with minutes to prepare.

Case Nexus loads every case into Claude's million-token context window. No RAG, no chunking. The AI sees everything at once.

Health Check. All 500 cases in a single Opus 4.6 prompt.

Claude uses 60,000 tokens of extended thinking — all streamed live.

The AI found Officer Rodriguez in four separate traffic stops — all with contested vehicle searches. A Fourth Amendment pattern invisible across 500 files.

Speedy trial deadlines, plea disparities, shared witnesses — the full risk profile.

(pause 5s)

Cascade Intelligence. Nine tools. Claude investigates autonomously.

Claude pulls case CR-2025-0089 for deeper investigation... now looking up the Georgia statute... searching for case law precedent...

A genuine agentic loop — Claude decides what matters, gets the information, builds strategy.

(pause 5s)

Three AI sessions chained. Prosecution. Defense. Judge.

The defense reads the prosecution's actual arguments and dismantles them.

80,000 tokens of chained reasoning across three perspectives.

Prosecution, defense, judicial — each with visible reasoning the attorney can challenge.

Claude also analyzes evidence using vision — dashcam footage, injuries, crime scenes.

Every Opus 4.6 capability — million-token context, extended thinking, autonomous tool use, 128K output, vision, chained reasoning — amplifying one attorney's judgment across 500 cases.

Voiceover Tips

  • Speak at 0.8x your natural pace. Judges watch at 1.5x.
  • Leave 1-second gaps between sentences. Feels slow recording. Perfect in playback.
  • Record in a closet or under a blanket fort. Clothes absorb echo. Phone Voice Memos is fine if the room is quiet.
  • No background music. Silence is better than bad music.
  • Don't apologize for loading time. Narrate what's happening instead.
  • Use concrete numbers. "275,000 tokens", "500 cases", "9 tools", "80,000 tokens" — specificity sells.

Speed Ramping Rules

Situation Speed Why
Cases loading in sidebar 2x Visual is clear at speed
Waiting for API response (blank screen) 3-4x Dead air kills
Token bar filling (before text appears) 2-3x Bar animation is clear at speed
Thinking text streaming 1x Judges need to read this
Results appearing 1x This is the payload — don't rush
Tool calls appearing one by one 1x Money shot — let each land
Scrolling through AI output 1x Let judges read key lines
Gap between adversarial phases Jump cut Use 0.3s fade-to-black

Backup Plans

Problem Solution
Health Check takes >90s Speed-ramp the wait. Caption: ~60 seconds of analysis
Cascade tool calls are slow Narrate what each tool does. Jump cut between calls
Adversarial Sim takes >3 min Show prosecution start → jump to defense results → show judge
API call fails mid-recording Stop, restart server, re-record just that take
Rodriguez pattern not found Re-run. It's planted in the data. If missed twice, use the take where it worked
Audio echo/noise Re-record voice only. Separate tracks means screen capture is safe

Export Settings

  • Resolution: 1920x1080 (1080p)
  • Codec: H.264
  • Frame rate: 30fps minimum (60fps if your screen capture is 60fps)
  • Bitrate: 10-15 Mbps (YouTube-quality)
  • Format: MP4
  • Total runtime: 2:45–2:55 (5-second buffer under 3:00 limit)

Final Checklist

  • Title card static/title_card.png used as opening frame
  • 5 takes recorded, best of each selected
  • Speed ramping applied (dead air fast, content 1x)
  • All 9 text overlays added
  • Jump cuts between adversarial phases with fade-to-black
  • Voiceover recorded separately and synced
  • Evidence analysis flash included (7 seconds)
  • Token bar visible during Health Check and Cascade
  • No console, no terminal, no errors visible
  • Mouse movements are deliberate (not jittery)
  • Total runtime ≤ 3:00
  • Exported at 1080p H.264
  • Uploaded to YouTube/Loom (unlisted OK)