JustTheGist

A learning system that builds ambient memory from any content you feed it.

JustTheGist is a persistent knowledge system that learns from every piece of content you analyze. Feed it any URL or file to learn from, or let it research a topic and build understanding from multiple sources. Everything you analyze gets stored in your personal knowledge base—building a richer, more interconnected understanding over time. Works with YouTube, podcasts, articles, PDFs, local audio/video. No code—just instruction files for Claude Code, Gemini CLI, Cursor, and more.

The Problem

You find an interesting-looking YouTube video, podcast, or article. It's 45 minutes long. You don't have time to watch/listen/read the whole thing just to find out if it has the information you need.

The Solution

Give the URL (or file) to your AI coding assistant. Tell it what you're hoping to learn. Get a detailed, tailored analysis in moments—complete with key takeaways, resources mentioned, notable quotes, and an honest assessment of whether the full content is worth your time.

Three Ways to Learn

Analyze (Learn from something specific you found) Give it a URL or file and what you're hoping to learn. Get a tailored analysis that builds your understanding.

Research (Explore a topic) Tell it what you want to learn about. JustTheGist searches, curates, and analyzes multiple sources—synthesizing findings into comprehensive understanding.

Recall (Access your accumulated knowledge) Query your personal knowledge base - everything you've analyzed before, semantically searchable for instant recall and connection-making.

Supported Content Types

Type	Source	Requirements
Online Videos	YouTube, Vimeo, Twitter/X, TikTok, Twitch, 1000+ sites	yt-dlp
Podcasts	YouTube, SoundCloud, Spotify (some), many podcast platforms	yt-dlp
Web Articles	Any URL	None (built-in)
PDF Documents	Local files	None (built-in)
Local Audio	.mp3, .wav, .m4a, .ogg, .flac	openai-whisper, ffmpeg
Local Video	.mp4, .mkv, .avi, .mov, .webm	openai-whisper, ffmpeg

Supported AI Tools

This repo includes instruction files for multiple AI coding assistants:

File	Tool
CLAUDE.md	Claude Code
GEMINI.md	Gemini CLI
AGENTS.md	OpenAI Codex
.cursorrules	Cursor
INSTRUCTIONS.md	Generic (copy to your tool's expected file)

Using a different tool? If your AI assistant reads from a project-level instruction file (like YOURTOOL.md), just copy any of the above and rename it. The instructions are tool-agnostic.

Prerequisites

• Python 3.8+
• An AI coding assistant (Claude Code, Gemini CLI, Cursor, etc.)

Quick Start

git clone https://github.com/vanclute/JustTheGist.git
cd JustTheGist
claude  # or gemini, cursor, etc.

That's it. Clone, open, use. Your AI assistant handles all setup automatically.

On first run, your AI assistant will:

1. Ask which content types you want to analyze
2. Automatically install only the dependencies you need
3. Save your preferences for next time

Then just give it a URL or file and say what you're looking for:

Here's a video about prompt engineering: https://youtube.com/watch?v=...
I want to learn practical tips I can use immediately.

Output

You get two things:

1. Immediate Summary - Key takeaways, resources, notable quotes, and a "worth your time?" assessment shown right in your terminal

2. Detailed Report - A comprehensive markdown file saved to docs/ with full analysis, all topics covered, and connections to your stated interests

Example Use Cases

Analyze Mode

• "I found this 2-hour conference talk. Does it cover anything about testing strategies?"
• "Here's a podcast episode. Extract any book recommendations they mention."
• "I have this PDF whitepaper. Summarize the key findings relevant to distributed systems."
• "Here's a YouTube tutorial. I just need to know how they handle authentication."
• "I recorded a meeting. What action items were assigned to the engineering team?"

Research Mode

• "Research the current state of AI coding assistants - find me 10 videos and tell me what the consensus is"
• "I want to learn about home automation. Find beginner-friendly content and summarize the key concepts."

Recall Mode

• "What did I learn about authentication patterns?" (searches your knowledge base)
• "Show me everything I've saved about React performance"

How It Works

No magic, no custom code. Just well-crafted instructions that tell your AI assistant:

1. Ask what you're hoping to learn
2. Detect the content type
3. Extract text (via yt-dlp for online video, Whisper for local audio/video, or direct fetch for articles/PDFs)
4. Analyze with your goals in mind
5. Generate a tailored report
6. Clean up temporary files

The AI does all the work. The instruction files just tell it how.

Cost Optimization

JustTheGist automatically optimizes for cost and speed by delegating tasks appropriately:

Task	Model Tier	Why
Extraction (transcripts, metadata)	Light (Haiku/Flash/GPT-4o-mini)	Mechanical work, no reasoning needed
Analysis & report writing	Standard (Sonnet/Pro/GPT-4o)	Requires judgment and synthesis

This can reduce costs by 10-20x for extraction tasks while maintaining analysis quality. Tools that don't support model switching simply use the default model for everything.

Knowledge Base

The Knowledge Base isn't a feature—it's the foundation. Every analysis automatically enriches your ambient memory, building a persistent "brain" from everything you analyze:

• Automatic Storage - Every analysis gets stored in your local vector database, no action needed
• Recall - Semantically search your accumulated knowledge anytime and see connections emerge
• Cite - Results include source attribution ("According to [Video] by [Channel]...")
• Compound Learning - Each new analysis draws from everything you've learned before, making future learning richer and more connected

Built on ChromaDB + sentence-transformers. Runs locally, no cloud required, works on CPU.

Additional dependencies (installed automatically if you enable this feature):

pip install chromadb sentence-transformers

The Knowledge Flywheel

The Knowledge Base creates a compounding effect similar to human learning:

New Content → Analyzed with prior context → Enriched synthesis stored → Informs next analysis
     ↑                                                                          │
     └──────────────────────────────────────────────────────────────────────────┘

What gets stored isn't raw transcripts - it's your interpreted understanding in context of everything already known. Over time:

• Connections between topics emerge organically
• Contradictions surface ("Source A says X, but Source B says Y")
• Consensus solidifies across multiple sources
• Understanding deepens with each piece of content analyzed

The more you learn, the better you learn. Each analysis is richer than the last.

Privacy

• All processing happens locally (except AI API calls to your chosen provider)
• Transcripts and reports stay on your machine
• No data sent anywhere except to the AI service you're already using

Contributing

Found a bug? Have a suggestion? PRs and issues welcome.

Ideas for contribution:

• Instruction files for additional AI tools
• Improved prompts for specific content types
• Better handling of edge cases

What Gets Installed (Automatically)

Curious what dependencies your AI assistant installs? Only what you need:

Capability	Package	Notes
YouTube videos	youtube-transcript-api + yt-dlp	Instant transcript fetch; yt-dlp for metadata
Other online video	yt-dlp	Supports 1000+ sites
Local audio/video	openai-whisper	Also requires ffmpeg (install via system package manager)
Web articles & PDFs	None	Built into your AI tool
Knowledge Base	chromadb, sentence-transformers	Local vector DB + embeddings

For local audio/video, you'll also need ffmpeg:

• Windows: winget install ffmpeg
• macOS: brew install ffmpeg
• Linux: apt install ffmpeg

License

MIT - Do whatever you want, just keep the license.

---

Stop watching. Start knowing.