System Prompts — Extracted from 30+ AI Coding Tools

What it is

This collection contains full system prompts extracted from 30+ AI coding tools including Claude Code, Cursor, Devin, Windsurf, Replit, v0, and others. Each prompt reveals the exact instructions that shape how these AI assistants behave, what constraints they follow, and what capabilities they expose.

The collection targets prompt engineers, AI researchers, and developers who want to understand how commercial AI tools are built and learn prompt design patterns from production systems.

How it saves time or tokens

Studying production system prompts saves weeks of trial-and-error in prompt engineering. You see proven patterns for instruction following, tool use formatting, safety guardrails, and output structuring used by the best AI products. Adapting these patterns to your own agents is faster than designing from scratch.

How to use

1. Browse the collection to find prompts from the tools you are interested in.
2. Study the structure: system context, behavioral constraints, tool definitions, and output formatting.
3. Adapt patterns to your own AI applications and agent configurations.

Example

# Common patterns found across AI tool system prompts:

1. Role definition

'You are an expert software engineer...'

2. Behavioral constraints

'Never modify files without user approval.'

'Always explain changes before making them.'

3. Tool use formatting

'When using tools, follow this exact JSON schema...'

4. Safety guardrails

'Do not execute destructive operations.'

'Ask for confirmation before deleting files.'

5. Output structure

'Respond with a brief summary followed by code blocks.'

Related on TokRepo

• Prompt Library -- browse reusable prompts for AI coding assistants
• AI Tools for Coding -- compare AI coding tools and their approaches

Common pitfalls

• System prompts change frequently. The extracted versions represent a point-in-time snapshot and may not match the current production version.
• Copying an entire system prompt verbatim is less useful than understanding the underlying patterns. Adapt the principles rather than the exact text.
• Some extracted prompts may be incomplete or from beta versions. Cross-reference with official documentation when available.