Free Claude skill
Let seven Claudes scroll past your video first, so you can fix the line where they leave.
It is a simulation with seven personas, not a prediction of real views.
Free with an account. No password: you get a login link by email, and it stays in your library.
A video file in Claude Code (it pulls a frame per second and the audio if ffmpeg is installed), or the script plus what is on screen, or just an idea, which becomes a first draft that gets tested. It also asks for the platform, who the video is for, what viewers should do afterwards and the planned length, but only for what is missing.
The script is numbered line by line with times and what is on screen. Real timings if it has them, otherwise an estimate that is marked as one. You see the beat list before anything is tested.
Seven Claudes play different people scrolling their feed. Each reads beat by beat and stops at the first beat where it would swipe away. None of them sees another's answer. Each gives the exact line, why, and the one change that would have kept it.
A text chart of how many of the seven reached each beat, with the biggest drop marked and the line quoted, then a short fix list tied to who raised each point. If three or more scrollers say the topic itself is the problem, it says the idea is the problem, not only the hook.
Ten new openings, each a different pattern, with a spoken line of at most twelve words and first frame text of at most six. The scrollers vote on them blind, and you get the votes and a recommended winner. A hook that needs footage you have not shown is marked and left out of the vote.
Outcome firstThe mistakeContrarian takeCall out the viewerShow the end firstBefore and after
The script is rewritten with the chosen hook: cut what does not move it forward, payoff earlier, on screen text for the sound off viewer, one short call to action. Your voice and facts stay. Then the seven scroll the new version and you see both drop maps side by side. At most one extra round, and only if you say yes.
Seven short verdicts, a drop map, a fix list, ten hooks with votes and a tighter script. This is the shape the verdicts and the map come in.
From each scroller, at most 70 words
The drop map
How many of the seven reached each beat.
The biggest drop is marked and the line is quoted, then the fix list.
No made up statistics about platforms or attention spans, and no promises about views. A hook never claims a number or result you have not shown. Where something is missing, it leaves a placeholder.
[ADD: the result you got]
It is honest, not kind: a boring opening can get seven swipes at beat one. It never makes a hook more extreme than the video can back up, and what only you can change (the footage, the first frame, the topic itself) is left to you.
scroll-test folder in ~/.claude/skills/ for every project, or in .claude/skills/ inside one project./scroll-test, with your script or video file.unzip ~/Downloads/scroll-test.zip -d ~/.claude/skills/Best here. The seven scrollers run as separate agents, so none of them sees another's answer.
There are no agents here, so Claude plays the seven one at a time. The blind test is simulated in one conversation.
Real viewers are quicker and less polite than seven personas. If you later bring your real retention graph, it compares it with the drop map and adjusts the next test.