Mike Hawk

opus 5.5 + google flow is CRACKED this whole JUBILEE style debate ad took about $1 of flow credits and is made entirely by claude code from scratch every talking shot came out of google flow. every frame came out of my ChatGPT. i never opened an editor. approved the frames, gave a few notes on pacing, done how it got built: - claude tore down a reference clip frame by frame - scraped how people actually argue about glp-1s online and wrote the script in their words ("you didn't earn it", "food noise") - turned real show frames into 3 completely new people, one image edit each - every talking shot is omni on google flow - pauses jump-cut out, captions, fine print and end card all done by claude the format works because nobody scrolls past two people arguing. RT + reply 'DEBATE' and i'll send you the full workflow (must follow so i can DM)

most people still don't comprehend how far AI video has come. no memory between clips used to mean every cut reset the scene, drifted, redrew itself. that's basically solved now if you build around it right this whole ad is one continuous WHITEBOARD, 9 chained clips, and the marker ink never moves, never redraws, never resets to a blank board. here's exactly how, so you can build your own version of this for any product the trick: OMNI flash has a hard 10 second limit per clip and zero memory of the clip before it. so you harvest the last frame of clip N, the board exactly as it looks at that instant, and feed it in as the starting frame of clip N+1. that's the entire mechanism. no cuts, no set, just a hand, a marker, and one continuous illusion the flow: - map your script to the arc: symptom, invisible cause, mechanism, solution, real product - write the VO as one continuous read, then split it into beats, ~20-25 words max per clip, one drawing action per clip - build a board map first, what's written and drawn on each row, so nothing collides as it fills up - generate clip 1, harvest its last frame, chain into clip 2, repeat - once the board gets dense (4-5 clips of ink in), "match the board" - stops being enough instruction. name every existing element explicitly, state a hard no-touch rule, and point at your target by position, not just by text, or the model will grab the wrong line - save the reveal for last: a real product photo beats a drawn one almost every time PROMPT: locked blocks, paste verbatim into every single clip STYLE: "Real documentary UGC footage, static locked camera on a dry-erase whiteboard, one real human hand and forearm only, never show a face or body. A single black marker is used for the entire clip, no color changes, no marker switching. Marker ink renders as genuine dry-erase marker, not a digital text or motion-graphic overlay. Handwriting is plain block capitals, legible at a glance. Icons stay minimal and schematic, one simple recognizable shape only. No captions, no subtitles, no on-screen text beyond what the hand physically writes or draws. ASPECT: 9:16 vertical." VOICE: "VOICEOVER (calm, warm, explanatory, American English, mid-30s):" opening clip, full prompt: "An 8-second locked-camera whiteboard explainer, vertical 9:16. Open on a blank white dry-erase board, static tripod shot, no camera movement, no zoom. 00:00-00:05: The hand writes your hook headline in black marker, plain block capitals, upper-left, then underlines it, natural handwriting speed, ink appears stroke by stroke, slight squeak. 00:05-00:08: The hand draws one small icon, upper-right. VOICEOVER: your hook line. AUDIO: marker squeak, quiet room tone underneath. No music. [STYLE LOCK]" every clip after that opens with "use the attached image, [clip N-1]'s final frame, as the opening frame, match the exact board state precisely" before describing the one new action for that beat