Someone just leaked 10,000+ lines of system prompts from the biggest AI tools—Cursor, Perplexity, v0, Lovable, Devin, and more. These aren't random configs. They're the exact instructions that make billion-dollar products work.
It's all in a public GitHub repo right now.
Quick note: This newsletter was written in 2 hours using the help of my Claude Code Newsletter Assistant - the same system I'm sharing with founding members of AI ContentLab OS.
If you're tired of spending 8+ hours on content creation, I'm building AI systems that maintain your voice while cutting writing time by 75%.
First 15 members join free forever, next 15 pay just $5/month - then I increase the price. Watch me build the complete content pipeline in real-time and get every tool I create. Join fast and become a founding member for free here.
The Leak: What's Actually In There
Who's Exposed: v0, Cursor, Manus, Same.dev, Lovable, Devin, Replit Agent, Windsurf Agent, VSCode Agent, Perplexity, Xcode, Trae AI, Cluely, Dia Browser, Spawn, Orchids.app, and others.
Why This Matters: You can now see:
How Cursor handles multi-file editing (hundreds of lines of rules)
Perplexity's exact search synthesis logic
v0's component generation patterns
Devin's task decomposition strategy
Reading these prompts is like getting the annotated source code for successful AI products. You see not just what they do, but how they think.
Immediate Value for Builders:
Copy proven conversation structures
Learn memory management techniques
Understand tone and personality tricks
See how different companies solve the same problems
The Missing Piece: Why Claude Code Wins
Here's what wasn't leaked: Claude Code, arguably the best AI coding agent available.
Developers prefer it over everything else. Same underlying Claude model, completely different results. Why?
Several developers have reverse-engineered it to find out.
A sneak peak into Claude Code’s prompt.
After decompiling 443,000 lines of JavaScript and intercepting API calls, they discovered something unexpected:
Claude Code's advantage is sophisticated prompt engineering, not a better model.
The evidence is striking:
Repetition = Reliability: The TodoWrite tool appears 5+ times in various forms throughout the prompt. Result: 95% execution rate. The lint command? Mentioned once. Result: 50% execution rate. The correlation is perfect.
System Reminders Fix Memory: After every task, Claude Code injects reminder blocks about the current goal. Without these, the agent forgets what you asked three messages ago. It's not your instructions—it's their memory.
Everything Is Natural Language: Task management, git workflows, code review—none of it is hardcoded. It's all prompt-based. Change the text, change the behavior. No code required.
Format = Function: XML tags like <example> and <task> aren't decoration. They create semantic structure the model can parse. ALL CAPS warnings actually work. Formatting improves task completion by 20%.
Model-Specific Optimization: These prompts only work well on Claude. Run them on GPT-4 or other models and performance drops 40%. Each model family needs different instructions for the same task.
The revelation: Anthropic built their competitive advantage with text. Not architecture. Not training. Just extremely well-crafted instructions, repeated strategically.
Your Playbook: Three Techniques to Implement Now
1. The 5x Rule
State critical requirements 5 different ways
Use variations: "Always," "NEVER forget," "IMPORTANT"
Place reminders after state changes
Result: 90%+ reliability on critical tasks
2. Structural Formatting
<task>
<description>What needs to be done</description>
<constraints>Specific requirements</constraints>
<examples>
<good>Correct approach</good>
<bad>What to avoid</bad>
</examples>
</task>
This structure improves comprehension significantly.
3. Progressive Reminders
Initial instruction: Set the goal
After first output: "Continue with [goal]"
After subtasks: "Remember main objective: [goal]"
Before completion: "Ensure [goal] is fully addressed"
The Real Lesson
The leaks prove something important: The gap between good and great AI isn't in the models (the difference becomes tighter with each release) - it's in how you talk to them.
Every company racing to build bigger models is missing the point. The winners are those who've mastered the language of instructing AI. The leaked prompts show you their vocabulary. Claude Code's techniques show you the grammar.
This isn't temporary. As models commoditize, prompt engineering becomes the differentiator. The company with better instructions wins, regardless of which model they use.
Your move: Download the leaked prompts. Study the patterns. Test the techniques. Build something better.
The "secret sauce" of every major AI company is now public.
What will you build with this knowledge?
Going Deeper: Check out the GitHub repository with prompts from 20+ companies. Developers have also published detailed reverse-engineering analyses and created visualization tools to understand Claude Code's architecture. Together, these provide a complete education in production AI systems.
That’s it for this week!
Luke