VideoCraft — Grounded Study Guides in Codex

Turn any technical YouTube video into a cited, interactive HTML study guide with Codex, timestamp evidence, original diagrams, search, and a quiz.

Comment

Inspiration

VideoCraft is a continuation of my previous SAP Community projects about chatbots, embeddings, OpenAI integration, and YouTube transcripts:

What I built

VideoCraft is a reusable Codex skill that turns a technical YouTube video into a portable, interactive HTML study guide. With Codex, I can generate Python scripts, run scripts and get insights about the created content, all in one. Codex helps with creating a plan, including testing and running until further instructions are required from my side. With the entire repo, Codex can explain what it does and can build and extend capabilities from my natural language prompt.

The workflow acquires the English captions and metadata, normalises the captions into stable evidence cue IDs, inspects the evidence, and creates a structured study.json. The result is rendered as a self-contained index.html.

The generated study guide includes:

  • a short explanation of the video;
  • chapters with timestamped evidence links;
  • original concept cards and a diagram;
  • a searchable interface;
  • an operational cheat sheet; and
  • a quiz with grounded explanations.

How Codex helped

I used Codex to read the repository instructions and the VideoCraft content contract, acquire and normalise the captions, inspect the evidence, author the study content, render the HTML, validate the evidence references, and run the Python test suite.

The repository includes the working VideoCraft skill, Python pipeline files, content contract, prompt, tests, generated example, and local usage instructions. The workflow can be copied locally and run through a Codex chat.

Challenges

The main challenge was making the study guide useful without copying and redistributing the full transcript. I used short timestamp ranges and original paraphrases, then checked that every referenced cue ID existed in the normalised evidence.

Another goal was portability. The final HTML is self-contained and can be opened locally without a server and API key.

What I learned

The project showed me that a good AI-generated study guide needs more than fluent text. It needs a clear evidence model, validation, and a simple way for the reader to check the source.

VideoCraft combines the ideas from my earlier chatbot and transcript projects with a stronger focus on traceability and practical learning.

Above all, I used Codex to generate the presentation, including voice narration, which helped me complete the project in one day due to limited time. I discovered the OpenAI Build Week a few days ago and first planned the idea. Then, with the idea in mind, I crafted the entire repo from prompts.

Built With

Share this project:

Updates