The remake, and the reasoning behind it
A smaller service with a stronger boundary.
The original SkillTube was a working Streamlit product with Google sign-in, Supabase usage records, a Stripe subscription, transcript extraction, and two Claude passes: one to identify candidate workflows and another to write the selected skills. It worked, but the machinery around acquiring and billing users became larger than the demonstrated demand.
This Lab edition keeps the useful sequence—validate, inspect, choose, generate—and removes the premature business. It uses a deterministic YouTube check and attempts captions before asking a model to inspect the video. Gemini is the practical fallback because it can understand a public YouTube URL directly; one fast model now handles both workflow discovery and skill writing. The generator follows the current portable SKILL.md convention, rather than the older Claude-specific metadata used by the first app.
Three analyses per week and three generated skills per video keep a no-account experiment predictable. The weekly limit is enforced by the server using an anonymous browser identifier and a one-way network signature. A second site-wide allowance stops distributed traffic after 24 analysis attempts in one day, videos are capped at 45 minutes when YouTube exposes their duration, and each signed analysis token can generate only one bundle. Those limits apply to direct API calls as well as the interface. They survive refreshes and ordinary cookie behavior, but no anonymous identity is literally impossible to evade; stronger person-level enforcement would require sign-in. ChatGPT sign-in could supply identity later, but a visitor's ChatGPT subscription cannot be used as this website's model API allowance.