I saved 40 reels and studied none of them, so Claude studied all of them

My admin has a studio page where I paste reels I want to remake. Forty references had piled up in there: twenty-eight Instagram reels, nine YouTube videos, two TikToks, one X post. The studio even has a proper study pipeline behind it, transcript in, anatomy out, shot list ready to film. Number of reels I had actually run through it: zero. A swipe file you never open is just a bookmark graveyard with better UI.

So I put Claude Code on my VPS and told it to study all of them. Not the thumbnails, the videos. It downloaded the thirty short-form ones with yt-dlp, cut each into a twelve-frame storyboard with ffmpeg, dense on the first three seconds because that is where the game is decided, and transcribed twenty-four of them with Whisper on the server's two shared vCPUs. Then it read every frame and every line, pulled the like and comment counts from the embed pages, and wrote me a playbook.

Two infrastructure notes before the findings, because they cost an hour. YouTube refuses downloads from a datacenter IP outright, so the nine YouTube references stayed at title and thumbnail level. And Instagram, famous for locking everything, serves its embed endpoint wide open: video, caption, like count, no login. The platform you expect to fight you is not the one that fights you.

Every winner stacks three hooks in frame one

Not one hook. Three, simultaneously, in the first frame, so the video works muted, unmuted, and at a glance. A text claim in five to nine words, bold white with a black stroke, top third of the frame. A face already mid-emotion, shock, disbelief, a hand on the head. And a spoken line that never repeats the text, it escalates it. The overlay says “RIP to every creator”, the voice says “I just made 72 UGC videos with Claude”. Text states, voice raises.

Most viewers scroll muted. The reels that won treat the text overlay as the actual video and the footage as its background.

Loss beats gain, and hooks are open source

The strongest pattern in the whole set is obsolescence. “MOTION editors are cooked”, “One guy just killed Higgsfield”, “Claude vient de tuer les designers”. Fear of being replaced outperforms every how-to phrasing in my saves, and it is not close: the cooked reel did 24,500 likes and 5,100 comments.

Better: three different accounts run the literal same text, word for word. Nobody in this niche treats a proven hook as property. It is infrastructure, you take it, you refill it with your own substance. That reframes the whole swipe file: I was collecting videos, I should have been collecting sentences.

The numbers follow the same logic. $67,291 a month, not six figures. 72 videos, not dozens. 1300€ in two hours. Specific reads as receipt, round reads as ad.

The skeleton under all of them

1. 01Seconds 0 to 3: the three-layer hook. Text claim, face mid-emotion, voice escalating.
2. 02Seconds 3 to 5: proof flash. The result shown before the method. The grid of finished videos, the payout screen, the site scrolling. It buys the next twenty seconds.
3. 03Seconds 5 to 25: the how, filmed as a screen story. A phone pointed at the laptop, never a clean screen capture, a finger or a pen pointing, and a caption every two to three seconds: go to claude.ai, click connectors, copy my prompt, let it cook.
4. 04Last ten seconds: the reveal, fast cuts of the output, one word on screen doing the celebrating.
5. 05Last three seconds: a one-word comment gate. GUIDE, HIGGS, OPEN, BANK. Comment the word, I send the link.

The phone-pointed-at-the-screen shot deserves its own sentence. Every one of these creators could record their screen properly. None of them do, because a filmed screen looks like a friend showing you something and a screen capture looks like an ad. The lo-fi is load-bearing.

And the comment gate is not an outro, it is the business model. All thirty short-form reels end on one. Comments explode distribution, the DM automation catches the lead, the freebie starts the funnel. One of the reels in my file collected 13,300 comments of people typing a single word at it.

The uncomfortable finding

The best performing reel in my entire swipe file, 48,700 likes, is a man in a car making an increasingly unhinged face for thirty-four seconds under five words of text: “I keep trying but I keep failing”. No tutorial, no screen, no product. Production value zero, emotion carried the whole thing.

I keep waiting until a video is polished enough. Nothing in the data supports the wait. The winners were shot on a phone in natural light the day the idea existed.

What changes for me

The pill text comes first now. Before any shot list, the five to nine words that would survive on a muted phone screen, and if they do not stop a scroll on their own the video does not exist yet. The playbook turned into a queue of hooks refilled with my own receipts, the miniature builds, the prompt pack, this exact server. And the eight best references go through the studio's proper anatomy pass next, so the swipe file finally does the job it was built for.

saved      40 references
studied    30 downloaded, 360 frames, 24 transcripts
finding    3-layer hook, proof flash, comment gate
shipped    1 playbook, 7 hooks queued
cost       0 new tools