This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady.
What I Built
You find an AI tutorial you want to try. Then come the pauses: find the repository, set up the tools, follow along, and figure out why your result looks different. Another useful video joins the "I'll try this later" collection.
I built SkillWalk for that moment.
It turns three reviewed YouTube tutorials into local workflows you can watch. Confirm the video, review the analysis, and approve execution. From there, it runs the supported steps, checks the result in a real browser, attempts bounded repairs when checks fail, and delivers the checked output without another required click. Stop remains available.
The main tutorial produces a PDF question-answering app. Presentation and frontend-design tutorials are also available.
You can paste your own YouTube link to check whether SkillWalk supports it. Currently, execution is available for three reviewed tutorials; other links are preview-only. SkillWalk uses local Gemma models, so tutorials designed around frontier Claude or OpenAI models may produce different results or require capabilities this version doesn't support.
Demo
Watch on YouTube. This animated promo illustrates SkillWalk's workflow.
The built-in document run needs no writing or upload. SkillWalk supplies a labelled synthetic PDF, opens the resulting app, and tests questions against its table and chart.
In the application, local model work, actual browser activity and the finished app are shown as distinct stages. Afterwards, you can ask your own questions, download the checked source, or use Try it yourself to follow the original tutorial independently.
SkillWalk runs locally; there is no hosted app.
Code
The repository includes the MIT-licensed application and local setup instructions.
How I Built It
The document workflow follows Nariman Codes' tutorial and its MIT-licensed source.
Local Gemma interprets the actual video captions, sampled frames and source materials, then produces configuration instructions. Reviewed adapters run the supported steps instead of executing commands taken from video text.
The app extracts PDF text, tables and images. Gemma 3 4B describes visual content and answers questions; EmbeddingGemma supports local retrieval. Node coordinates the workflow, Streamlit provides the resulting app, and Playwright checks it inside an isolated local Docker runtime.
The result is a working local app where you can ask questions about a PDF, see the retrieved source passages, and try your own document after the automatic run finishes. The checked source is available to download so you can inspect or adapt the workflow.
Why Does Open Innovation Matter?
SkillWalk builds on reusable source code and runs its model inference on your own computer. It needs no paid model API or model-service key, and PDF contents are not sent to a hosted inference provider. Instructions and code are available to inspect and adapt.
That makes repeated attempts possible without a per-request model bill. Hardware, electricity and setup still have costs. YouTube and initial downloads need internet access.
There are tradeoffs: the document video demonstrates a larger Gemma model, while the other two tutorials demonstrate Claude. This local adaptation does not promise the same appearance or answer quality.
The Lazy Payoff
Instead of pausing, rewinding and following every step yourself, give SkillWalk a supported YouTube tutorial. It analyzes the video and prepares a plan. Approve it, then sit back and watch the workflow run—from creating the result to testing it and making supported repairs.
When it finishes, explore what it made or download the source to try it yourself.
Prize Categories
Best Use of Gemma: local Gemma performs source interpretation and configuration, image/table description and document answers. EmbeddingGemma supports retrieval. These are working parts of the product, rather than an added summary.
Top comments (0)