Handoff — Reverse-Engineer a Video → Agent Opus Prompt + Storyboards

Date: 2026-06-27 Project: Agent Opus (the video-creation engine — feeds Why Dachshunds + doTERRA/Monthly Drop) Status: TABLED — figured out the tooling, not yet tested. Pick up when ready.

---

What Sondra wants

Two related abilities, both seen in the Jun 26 Agent Opus office hours (Dr. Abdullah's workflow — meeting is in Granola/IntelOS):

1. Reverse-engineer an existing video → a forensic, second-by-second prompt that recreates

it (every visual, color, camera move, transition, sound effect + audio sync). His prompts run 50-60 pages. He feeds this long prompt into Agent Opus Story Mode.

2. Storyboard for a NEW idea → one multi-panel storyboard image fed into Agent Opus as a

style/structure reference, alongside the long prompt. Dr. Abdullah made his storyboard with Higgsfield — Sondra does NOT want Higgsfield (it's Sara's account).

His full move = one storyboard image + one very long prompt → Story Mode.

The answer (already worked out)

No new program needed. Everything's already installed/owned — confirmed 2026-06-27:

| Piece | Status | Role | |---|---|---| | ffmpeg (v8.1.1) | installed | chop video into frames + pull audio | | yt-dlp | installed | download video from a URL | | claude-video / /watch skill | installed | download → frames → transcript → hand to Claude (vision) | | Whisper transcription | configured (~/.config/watch/.env) | reads the audio for sound/voiceover | | Gemini image gen (gemini-3-pro-image-preview) | owned, already used for thumbnails + kids' book | the Higgsfield replacement — generates the storyboard grid image | | Claude (me) | — | writes the shot list + storyboard + long prompt |

Honest caveat: "frame by frame" really = smart sampling ~1 frame/sec (skill caps ~60-100 frames). Enough to reconstruct scenes/color/motion/timing. Literal 30fps isn't practical and isn't what he does either.

Division of labor (the reassurance — Sondra doesn't write anything technical)

• Sondra brings: the idea in one sentence ("do a dachshund video about ___"), or a video to copy.
• Claude does: shot list → storyboard image (Gemini) → long Agent Opus Story Mode prompt.
• Sondra does: glance, approve/tweak, paste into Agent Opus. (She does NOT write shot lists.)

Also coming

Jake's office-hours recap doc (due Jun 26/27) reportedly includes a Claude skill for prompt creation + a ChatGPT GPT for visual storyboard generation. Grab both when they land and compare against the Gemini path.

Next step when resumed

1. Test on ONE short video/idea (under ~2 min) — prove both paths:

• Reverse-engineer: feed a video → get the forensic recreation prompt.
• Storyboard: feed an idea → shot list → Gemini storyboard grid → long prompt.

1. If it lands → package into a repeatable skill ("reverse-engineer-video" / video→prompt)

so it's one command every time.

Related

• Agent Opus MCP server just launched (Jun 26 office hours) — bulk video (10/5-min window),

named anchor images, plan-then-generate. Separate thread; could replace the HAR-hack CLI.

• Quality fix pushed this week (color/hex adherence) — rerun any project from last 7-10 days

that had color issues.

• Existing related skills: agent-opus-prompt-builder (TACKL, script→prompt),

wellness-agent-opus-prompt-builder (doTERRA script→prompt), video-prompt-builder, claude-video. None do video→prompt reverse-engineering yet — that's the gap to build.