Video is a rich source of procedural knowledge and a poor format for an agent that needs to apply the same procedure tomorrow. SkillTube attempted to compile that knowledge into a reusable SKILL.md rather than repeatedly asking a model to rediscover it.
The original was a Streamlit application that processed a YouTube source into structured, reusable instructions for an AI-agent skill. A new no-account Lab edition now preserves the useful workflow in a smaller public service. Its purpose is to convert long-form video knowledge into a structured SKILL.md that an AI agent can reuse.
The question behind SkillTube
The distinction is between summarization and operationalization. A useful skill must preserve the steps, conditions, tools, and boundaries an agent needs to act, not merely produce a shorter description of the video. It also needs a human checkpoint: a model can propose candidate workflows, but the visitor should decide which procedures deserve to become durable instructions. Agent users, educators, and developers may benefit. Video creators and rights holders are affected by transcription, transformation, and redistribution choices.
How SkillTube took shape
The recovered Streamlit source validates a YouTube identifier, attempts manual and generated captions, asks Claude to identify repeatable workflows, lets the user select them, and asks Claude again to write downloadable skills. Google OAuth, Supabase weekly usage, Stripe, and Render surrounded that core. Contrary to Josiah's recollection of one iteration, the recovered source contains no Gemini fallback. The Lab remake preserves the two-stage approval flow while using deterministic caption extraction first and Gemini's direct public-YouTube understanding as the fallback. Josiah identified the video-to-agent-knowledge opportunity, directed the original product, evaluated its failed acquisition model, and chose to revive the capability as a free public experiment with explicit cost and review boundaries.
What the evidence supports
The original application worked technically but did not establish a viable user-acquisition path. Its source was recovered and audited, and the rebuilt Lab interface and server workflow are now part of the published site. The capability survived the first product model. The remake deliberately avoids sign-in and payment while limiting anonymous visitors to three weekly analyses and three generated skills per video. That boundary is sufficient for a public experiment, not a claim of unbreakable identity: refreshes do not reset the server record, but only authentication could make a person reliably persistent across browsers and networks. SkillTube now tests the underlying value before rebuilding the acquisition machinery that once obscured it.
Output quality depends on the source and model, generated instructions require review, caption extraction is fragile, Gemini's direct YouTube feature is a preview, anonymous rate limits can be evaded by changing identity and network, and source rights must be respected. A small public beta will test generated skills across varied tutorial types, expose recurring failure modes, and clarify whether SkillTube should remain a focused tool or become an ingestion path inside AgentWorkbench.