Deconstructing the system prompts behind top AI tools

What it does

The repository gathers actual system prompts from tools like ChatGPT, Claude, Vercel v0, Manus, same.new, Cline, and Bolt.new, then analyzes them in a lengthy README. It treats these prompts as case studies to identify recurring patterns—clear role definition, structured instructions, explicit tool schemas, and safety protocols—that keep agentic AI from going off the rails. Think of it as a field guide to the instruction manuals vendors never publish.

The interesting bit

Instead of offering generic prompt-engineering tips, the project directly quotes and compares how different products format their internal constitutions, from ChatGPT’s TypeScript tool schemas to Manus’s XML-like capability tags. The analysis reveals that the best prompts read less like creative writing and more like strict API documentation written for a very literal employee.

Key highlights

• Side-by-side breakdowns of prompts from ChatGPT (4.5/4o), Claude Sonnet 3.7, Vercel v0, Manus, same.new, Cline, Bolt.new, and others
• Identifies eight core principles, including explicit tool integration, step-by-step reasoning loops, and refusal protocols
• Shows how vendors structure instructions using Markdown headings, custom XML tags, or embedded TypeScript schemas
• Highlights architectural differences: some prompts embed tool definitions inline, while others reference external JSON schemas

Verdict

Worth bookmarking if you are building an AI agent and need to write a system prompt that actually constrains model behavior. Skip it if you are looking for a drop-in library or framework—this is purely reference material and analysis.