Course Transcriber
A client needed course material to become a useful local reference, rather than hours of video scattered across learning platforms. I built Course Transcriber to turn accessible course libraries into organized Markdown transcripts, with an AI agent handling the sites and the tool handling repetitive acquisition and transcription.
Knowing what belongs in the library
Finding a video is only part of the job. A course site can contain repeated embeds, trailers, attachments, multiple lesson layouts and lessons behind different navigation paths. I made a reviewed site manifest the basis of the workflow: the agent inspects the site and decides what belongs, then the tool processes that explicit set.
That gives completion a useful meaning. The output can be checked against the reviewed lessons, rather than declaring success because some files appeared in a folder. Known duplicates and irrelevant media are excluded deliberately.
Let the agent investigate; let the tool repeat
The agent handles login, representative-page inspection and unusual media. The CLI maps known site structures, acquires validated audio, runs local transcription batches and records progress. Completed work survives interruptions so a long run does not have to begin again.
I worked through real course-platform differences and built the same overall workflow for macOS and Windows. Windows setup includes a pinned local speech runtime, with CUDA support where an NVIDIA driver is available. The end result is an ordinary transcript folder, useful outside the tool that created it.
The important output is the course material in a form the client can search, read and reuse. The browser automation and transcription machinery support that outcome; they do not need to become another interface the client has to learn.
A folder you can keep.