Gitingest — Convert GitHub repos into a clean, LLM-friendly format

This is not sponsored. I made it for two reasons:

1. Bring your attention to gitingest (which is a neat little tool).

2. Show you how to chat with codebases using small, local models (like gemma3).

Check out the video on YouTube

Course Update: Audio

Lesson audio is here! 🎉

I’ve added AI narrations for all 96 lessons.

I even transcribed the math and code snippets.

Have a listen and let me know what you think.

Topics from this week’s video

• Gitingest Overview

  • Tool types: web app, CLI tool, and client library

  • Converts GitHub repositories into LLM-friendly text files

  • Specialized utility to help language models understand codebases

• Video Purpose

  • Showcase Gitingest's functionality

  • Demonstrate integration with local language models

  • Share use cases and terminal workflows

• Web Application Walkthrough

  • Ingests a Git repo and generates:

    • Summary (repo name, file count, token count)

    • Directory tree (Git-like file structure)

    • Flattened text digest for LLM input

  • Includes options to filter contents (e.g., only source folder)

  • Example repo used: neo-brutalism-styled component library

• CLI Tool Installation and Setup

  • Recommended install via pipx for isolated environments

    • Demonstrated use of which and symlink tracing

    • Pipx installs to virtual env and links executable

  • Local ingestion example: Discord bot repo

    • Ingests current directory

    • Warns of including sensitive files like credentials

    • Demonstrates excluding specific files with -e flag

• Using Gitingest Output with Local LLMs

  • Runs Gemma 3 model locally via llm CLI tool

    • Uses 4B and 12B parameter versions

    • Prompts include “Explain like I’m 5” style questions

    • Shows system GPU usage and model performance

  • Highlights pros and cons:

    • Free local inference

    • Occasional infinite loops (e.g., in Gemma 3)

    • Performance monitored with nvtop

• Improving Workflow

  • Demonstrates aliasing repetitive commands (dq)

  • Uses digest + prompt combo for streamlined querying

  • Real-time demo of model response latency

• Remote Repository Example

  • Analyzes requests Python library from GitHub

    • Full ingestion = too many tokens (2.6M)

    • Solution: include only /source directory (-i source)

    • Output digest becomes manageable (~42K tokens)

• Advanced Prompting and Iteration

  • Example use case: extend requests to support JS rendering

    • Model suggests using Playwright

    • Promotes structured prompt chaining via llm -c (conversation mode)

    • Notes caveats: must let prior completions finish to maintain context

---

If you believe your product or service can fulfill a true need, it is your moral obligation to sell it.

Zig Ziglar