Repo Teacher - a step towards automated developer education

Dev tool startups GTM through developer education and make sample projects to teach developers. But it's super time consuming. Repo teacher teaches learner devs so product engineers don't have to.

Comment

Inspiration

The problem: AI dev tool companies understand that education through sample projects and tutorials is the best way to go to market (e.g Vercel, LangChain, TogetherAI, etc..), but many teams don't have capacity to make good explainer or walkthrough videos. This tool takes a repo and walks the learner through it step-by-step, reducing the need for video production or technical writing.

I launched a platform for genAI developer tooling companies to go-to-market through educational video content (think Coursera, or DeepLearning.ai). My insight is: dev tool companies can dominate a new market if they can educate it through feature-rich example educational code projects (like Vercel, LangChain, CrewAI, Supabase had already done). However, many of my customers don't have capacity to upload videos regularly because their teams are smaller (<10 people). So to fill this gap, I built Repo Teacher so any engineer can create an educational sample repo, and it will walk the learner through the repo instead of the engineer. This will increase my customers' output and get more value out of my content platform.

What it does

It gives you step-by-step instructions to build any feature that is part of a GitHub repository that you give it (e.g given the Vercel Chatbot repo, which is very dense in features but you just wanted to know how to build messages persistence, you can drop the URL of the repo in here and look up the feature you want to learn. Then it'll give you step-by-step guide on how to build messages persistence)

How we built it

Tech stack: Next.js, Vercel, Vercel's AI SDK, OpenAI o3-mini, Gemini 2.0 Flash

  1. UI prototyped on v0
  2. Repository extraction using Repomix
  3. Gemini model looks through the repo and extracts a list of features available
  4. OpenAI o3-mini (with strict structured ouput) creates a list of steps (JSON objects) that a learner can go through to learn the feature.

Challenges we ran into

  1. The context of a large codebase easily exceed an LLM's context window, so larger codebases' performance may not be optimal.
  2. Getting structured outputs from non-thinking models like Gemini-2.0-Pro was unreliable, so models
  3. Need better evals for walkthrough steps so they're more learner-friendly and shorter
  4. Vercel's serverless functions timeout sometimes because o3-mini (with low reasoning effort) takes a long time.

Accomplishments that we're proud of

  1. The user flow works end-to-end
  2. At least 4 of my coding students using it right now and burning my OpenAI credits since last night.

What we learned

  1. Quick prototyping on v0 -> build it out with a comfortable stack like Vercel + AI SDK -> deploy to Vercel was key to getting this done.
  2. Better evals means better outputs

What's next for Repo Teacher

I'll look into automating the educational repo creation with an AI Agent that goes through the developer's docs and create sample use case apps (then it'll cover the entire developer education flow, end-to-end, giving small teams the power of larger companies with dedicated learning departments)

Built With

  • gemini-2.0-pro
  • next.js
  • openai-o3-mini
  • vercel
  • vercel's-ai-sdk
Share this project:

Updates