See You Tomorrow

Aaron at 33 — a film by his favorite twin brother

3:46 · 1080p · July 7, 2026

The brief

My twin brother Aaron turned 33 on July 7, 2026 — our birthday. In August my family makes aliyah to Raanana, Israel, so this was the last birthday we'd spend as neighbors. I gave Claude (Anthropic's Fable 5, running in Claude Code) a folder and one instruction: make a 1–5 minute film celebrating Aaron, mix real footage with AI, use Whisper on the WhatsApp voice notes, budget $10, make the viewer cry.

Everything below — research, script, shot selection, AI generation, edit, and mix — was done by the agent in a single session, in about two hours.

The raw material

  • A 6 GB photo/video archive (457 files) — including photos of physical photo-album pages from the 1990s
  • A full WhatsApp chat export: 9,593 messages between the two of us from January to July 2026, plus 1,900+ photos, videos, and voice notes
  • Whisper (faster-whisper, running locally) transcribed every voice note — which is how the film found its cold open: a real voice note that begins “Hello good sir, it's your favorite twin brother.”

The agent built contact sheets of all 309 photos and 130 videos, read them visually, and cast the film from them — the scanned photo of us as toddlers with two birthday cakes became the emotional anchor. Data-mining the chat produced the stats card: 120 “good sir”s, 98 “love”s, 91 conversations about Claude. And the climax was sitting in the archive all along — a real exchange from Purim: “Plz come our last Purim together” → “see you tomorrow and next year in raanana inshallah” → “Amen!”

The AI shots (Gemini Omni Flash)

Six video generations via gemini-omni-flash-preview (the Interactions API). Five text-to-video establishing shots and one image-to-video resurrection: the 1995 two-cakes photograph, animated so the toddlers actually blow out their candles.

Sample prompt (the opening shot):

Vintage 16mm film, summer morning, July 1993. Toronto skyline with the CN Tower, soft golden haze, slow cinematic push-in toward a hospital building window. Film grain, warm nostalgic color, gentle ambient city sounds, soft hopeful piano. Single continuous shot, no scene cuts, no dialogue, no text on screen.

And the image-to-video prompt for the two-cakes photo:

Bring this old 1990s family photograph gently to life: the candle flames flicker, the two toddlers lean in and blow out the candles on their two birthday cakes, the adults smile and clap. Subtle, warm, realistic home-video motion. Keep the vintage 90s film look, colors and everyone's faces exactly the same. Single continuous shot, no scene cuts.

The nursery shot — two bassinets labeled MILLER A and MILLER B, the babies reaching toward each other — came back perfect on the first try.

Voice, music, and the mix

  • Narration: ~340 words written from my point of view, voiced with minimax/speech-02-hd on Replicate (7 segments, ~2.9s average generation time)
  • Score: two 30-second cues from meta/musicgen (stereo-large) — nostalgic piano for the childhood acts, hopeful swell for the ending — looped with crossfades into a full bed
  • Real audio: Aaron's actual voice notes, a playground vlog (“one, two, three, go!”), and my daughters shouting “Hi, Uncle!” over the credits
  • Mix: a single ffmpeg filtergraph — music side-chain compressed (ducked) under every voice, limited, and muxed against the picture

The edit pipeline

No editing software — the whole cut is Python + ffmpeg, fully scripted and resumable:

  1. Every photo becomes a 1920×1080 “plate” (blur-fill background) with a Ken Burns zoom via zoompan
  2. Typographic cards (title, WhatsApp bubbles, stats, end cards) rendered with PIL — Georgia for the emotional cards, Avenir Next for the data
  3. All 44 shots render to uniform silent MP4s, concat-muxed into picture lock
  4. The audio mix is computed from the same timeline data, so every voiceover and voice note lands on its intended frame

What it cost

Item Usage Approx. cost
Gemini Omni Flash (video gen)6 × 10s clips, 720p~$1–2 (preview pricing)
MiniMax speech-02-hd (Replicate)7 segments, ~2,200 chars~$0.15
MusicGen stereo-large (Replicate)2 runs, ~88s GPU each~$0.25
Whisper transcription15 voice notes/audio files$0 (local)
Totalvs. a $10 budget≈ $2

Why it works

The AI shots are connective tissue, not the substance. Every emotional beat in the film is real: the scanned photographs, the voice notes, the chat messages, the statistics of an actual friendship. The models just gave the memories a camera.

Happy birthday, Aaron. Next year in Raanana. Amen.

● The Workshop · weekly

Everything I build, in one email a week.

Torah tools, daf games and AI experiments, with the story behind each one. Sign up and this week’s issue arrives within the hour.