Hook Generator

Ask any model for ten hooks and you get ten hooks. They scan fine, and they are invented. The model returns the statistical middle of every opener it has read, which is why everyone's first line now sounds the same. I shipped a version of this skill that worked exactly that way, and the verdict I got back was one word: useless. So I inverted it. hook-generator carries 64 hook templates, each lifted from a real published post and each carrying the URL it came from. Eleven are the first 3 seconds of real short videos. I read the best and the worst video of 12 creators frame by frame: 8 of the 12 worst opened on the creator themselves, and 1 of the 12 best did. Your agent matches on platform and angle, installs the templates, and writes your topic into shapes that already ran. Then it gates itself: six scripts check length against per-platform ceilings, lint for AI-slop with a non-overridable BPR and CLP gate for regulated products, demand real variety, fail a launch headline that sells how the thing was built, and check a video hook's words, on-screen text and first frame. Eighteen templates carry a verifiable metric; the rest say pattern-level rather than carrying numbers nobody measured. It writes the opener and nothing else.