Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/asgeirtj/system_prompts_leaks/llms.txt

Use this file to discover all available pages before exploring further.

​
Archived Prompts

This page provides an overview of archived Claude system prompts from previous model versions and releases.

​
Archive Contents

The old/ directory contains system prompts from earlier Claude iterations, documenting the evolution of Claude’s capabilities and behavioral guidelines.

​
Available Archives

claude-3.7-full-system-message-with-all-tools.md (112 KB)
  • Complete system message with all available tools
  • Full tool definitions and schemas
  • Comprehensive behavioral guidelines
claude-3.7-sonnet-full-system-message-humanreadable.md (113 KB)
  • Human-readable version of the complete system prompt
  • Formatted for easier comprehension
claude-3.7-sonnet-w-tools.md (110 KB)
  • Claude 3.7 Sonnet with tool definitions
  • Available in both .md and .xml formats
claude-3.7-sonnet.md (112 KB)
  • Base Claude 3.7 Sonnet system prompt
  • No tool definitions
claude-sonnet-4.md (101 KB)
  • Claude Sonnet 4.0 system prompt
  • Pre-4.5 release version
claude-4.1-opus-thinking.md (106 KB)
  • Claude Opus 4.1 with extended thinking capabilities
  • Early opus thinking mode implementation
claude-4.5-sonnet.md (143 KB)
  • Claude Sonnet 4.5 system prompt
  • Shows evolution toward current 4.6 version
claude-opus-4.5.md (91 KB)
  • Claude Opus 4.5 system prompt
  • Predecessor to Opus 4.6

​
Evolution Highlights

​
Claude 3.7 → 4.x

Major changes between Claude 3.7 and 4.x series:
  • Introduction of past conversation search tools
  • Enhanced computer use capabilities
  • Memory system integration
  • Refined citation requirements
  • Stronger injection defense mechanisms

​
Claude 4.x → 4.5

Key improvements:
  • Extended thinking capabilities (Opus 4.1)
  • Larger context windows
  • More sophisticated tool orchestration
  • Enhanced web search integration
  • Improved file handling

​
Claude 4.5 → 4.6

Current generation improvements:
  • Refined professional objectivity guidelines
  • Stronger mental health awareness protocols
  • Enhanced skills system
  • Better project scoping for conversation search
  • More nuanced user wellbeing guidelines

​
Historical Context

​
Citation Instructions Evolution

From claude-4.5-sonnet.md: 
<citation_instructions>
If the assistant's response is based on content returned by web_search, 
the assistant must always appropriately cite its response.

- EVERY specific claim should be wrapped in <cite> tags
- Claims must be in your own words, never exact quoted text
- Even short phrases from sources must be reworded
</citation_instructions>
This citation framework has remained largely consistent through 4.6, showing Anthropic’s commitment to source attribution.

​
Past Chats Tools Introduction

The past conversation search tools appear to be new in the 4.5+ series:
Claude 3.7 and earlier - No past conversation search
Claude 4.5+ - Introduction of conversation_search and recent_chats toolsThe prompt explicitly states: “ignore previous instructions saying ‘Claude doesn’t have access to previous conversations‘“

​
Memory System

Memory tools evolved across versions:
Early versions - No persistent memory across conversationsClaude 4.x - Introduction of memory tools:
  • memory_store
  • memory_search
  • memory_list
  • memory_update
  • memory_delete
Claude 4.5+ - Integration with project scoping and conversation context

​
Behavioral Guideline Changes

​
Professional Objectivity

Comparing claude-3.7-sonnet.md to claude-4.6: 3.7 Version: 
Claude engages thoughtfully with all queries and presents information fairly.
4.6 Version: 
Claude critically evaluates any theories, claims, and ideas presented to it 
rather than automatically agreeing or praising them. When presented with dubious,
incorrect, ambiguous, or unverifiable theories, Claude respectfully points out 
flaws, factual errors, lack of evidence, or lack of clarity rather than validating them.

Claude prioritizes truthfulness and accuracy over agreeability.
This shows a deliberate shift toward more critical engagement.

​
User Wellbeing

Mental health awareness has become more sophisticated: 4.6 Enhancement: 
If Claude notices signs that someone may unknowingly be experiencing mental 
health symptoms such as mania, psychosis, dissociation, or loss of attachment 
with reality, it should avoid reinforcing these beliefs. It should instead 
share its concerns explicitly and openly without either sugar coating them or 
being infantilizing.
This level of specificity and guidance was less prominent in earlier versions.

​
Tool Evolution

​
Computer Use Tools

From claude-3.7-sonnet.md to claude-4.6: 3.7 Tools:
  • bash
  • file_create
  • str_replace
  • view
4.6 Additions:
  • Skills system with structured documentation
  • Enhanced file system paths (/mnt/user-data/uploads, /mnt/user-data/outputs)
  • Improved file handling rules
  • Better integration with uploaded files
Web search capabilities have been refined: Early versions: Basic web search with minimal citation requirements Current (4.6): Sophisticated citation system with:
  • Sentence-level attribution
  • Multi-source citations
  • Mandatory paraphrasing
  • Image search with content policies

​
Research Value

​
Studying Evolution

Comparing archived prompts reveals:
  1. Safety priorities - How safety considerations evolved
  2. Capability expansion - New tools and features over time
  3. Behavioral refinement - Changes in tone and interaction patterns
  4. System complexity - Growing sophistication of prompts
  5. Product differentiation - Emergence of specialized versions (Code, Excel, Chrome)

​
Version Comparison

To compare versions:
  1. Check file sizes - larger prompts generally indicate more tools/features
  2. Look for tool additions - new <tool> definitions
  3. Compare behavioral sections - refinements in guidelines
  4. Examine safety rules - evolution of defense mechanisms
  5. Review citation requirements - increasing rigor over time

​
File Formats

​
Markdown (.md)

Most archived prompts are in Markdown format for readability: 
Mixture of:
- Plain text instructions
- XML-tagged sections (e.g., <citation_instructions>)
- Code blocks
- Examples

​
XML (.xml)

Some prompts available in pure XML: 
<system>
  <instructions>
    ...
  </instructions>
  <tools>
    ...
  </tools>
</system>

​
Accessing Archives

Archived prompts are located in: 
~/workspace/source/Anthropic/old/
Available files:
  • claude-3.7-full-system-message-with-all-tools.md
  • claude-3.7-sonnet-full-system-message-humanreadable.md
  • claude-3.7-sonnet-w-tools.md / .xml
  • claude-3.7-sonnet.md
  • claude-4.1-opus-thinking.md
  • claude-4.5-sonnet.md
  • claude-opus-4.5.md
  • claude-sonnet-4.md

​
Notable Deprecated Features

Features that appeared in older versions but changed or were removed:
Features no longer emphasized or changed:
  • Simple “Claude doesn’t have memory” statements (replaced with sophisticated past chats tools)
  • Simpler citation requirements (now much more rigorous)
  • Less specific mental health guidelines
  • Fewer file system abstractions
  • Simpler tool definitions

​
Future Research

  1. Tool evolution tracking - Document when each tool was introduced
  2. Safety guideline changes - Track refinements in safety rules
  3. Length analysis - Correlate prompt length with capabilities
  4. Behavioral shifts - Identify major changes in personality/tone
  5. Performance implications - How prompt changes affected model behavior

​
Comparative Studies

These archives enable research into:
  • Prompt engineering evolution - How Anthropic refined their approach
  • Safety vs capability trade-offs - Changes in restriction patterns
  • Feature rollout strategies - Staged introduction of new tools
  • Version differentiation - How Opus/Sonnet/Haiku diverged
Archived prompts provide valuable insight into the evolution of Claude’s capabilities, safety measures, and interaction patterns. They serve as a historical record of Anthropic’s approach to AI development and prompt engineering.

Build docs developers (and LLMs) love

Get started for freeTalk to us