Skip to content
gentic.news — AI News Intelligence Platform
Connecting to the Living Graph…

Listen to today's AI briefing

Daily podcast — 5 min, AI-narrated summary of top stories

A developer's terminal window showing git log output with commit messages referencing Claude system prompt changes…
Open SourceScore: 100

How to Use Git History to Analyze Claude's System Prompt Evolution

A new tool converts Anthropic's official system prompt history into a git timeline, enabling developers to analyze prompt changes with standard version control commands.

·Apr 18, 2026·3 min read··441 views·AI-Generated·Report error
Share:
Source: simonwillison.netvia simon_willison, medium_claudeWidely Reported
TL;DR

Simon Willison's tool transforms Anthropic's published system prompts into a git repository, letting you use `git diff` and `git blame` to track changes between Claude versions.

The Technique — Git History for System Prompts

Complete Guide to setting up Git Flow in Claude Code | by Daniel Avila ...

Simon Willison has created a tool that transforms Anthropic's published system prompt documentation into a git repository. Anthropic maintains a public Markdown file containing all system prompts for Claude chat models across versions. Willison used Claude Code to parse this monolithic document and create separate files for each model and model family, with fake git commit dates corresponding to when each prompt version was released.

Why It Works — Version Control as Analysis Tool

By structuring the prompt history as a git repository, developers can use familiar version control commands to analyze changes:

  • git log --oneline shows the chronological evolution of prompts
  • git diff <commit1> <commit2> reveals exact changes between versions
  • git blame attributes specific prompt sections to particular releases
  • GitHub's web interface provides visual diffs and commit browsing

This approach eliminates manual parsing of Anthropic's documentation and provides a structured way to understand how Claude's "personality" and capabilities have evolved through system prompt changes.

How To Apply It — Clone and Explore

You can immediately start analyzing Claude's prompt evolution:

# Clone the repository
git clone https://github.com/simonw/research
cd research/extract-system-prompts

# View the commit history
git log --oneline --graph

# Compare two specific versions
git diff claude-3.5-sonnet-20241022 claude-3.5-sonnet-20250220

# See what changed in a specific file over time
git log -p --follow prompts/claude-3.5-sonnet.md

Willison used this tool to write detailed notes comparing Opus 4.6 and 4.7, identifying specific additions and removals in the system prompt that might explain behavioral changes.

Practical Applications for Claude Code Users

How to Use Git Worktrees in Claude Code: Parallel AI Development Guide ...

  1. Debugging Behavior Changes: When Claude Code behaves differently after an update, check if system prompt changes explain it
  2. Understanding Model Strengths: Different model families have different system prompts—compare them to understand specialization
  3. Prompt Engineering Insights: See how Anthropic engineers structure effective system prompts
  4. Version-Specific Optimization: Tailor your own prompts based on what each version's system prompt emphasizes

Try It Now — Quick Analysis Commands

Here are specific commands to get immediate insights:

# See all Opus versions and their release dates
cd extract-system-prompts
git log --oneline prompts/claude-opus.md | head -20

# What changed between the last two Sonnet versions?
# First, get the last two commits for Sonnet
git log --oneline prompts/claude-3.5-sonnet.md | head -2
# Then diff them (replace with actual commit hashes)
git diff abc123 def456 prompts/claude-3.5-sonnet.md

# Find when a specific phrase was added to any prompt
git log --all -p --grep="chain of thought" | head -100

The repository structure includes separate files for each model family (Opus, Sonnet, Haiku) and individual model versions, making targeted analysis straightforward.

Integrating with Your Claude Code Workflow

While this tool doesn't directly modify how Claude Code works, it provides valuable context for:

  • Writing better CLAUDE.md files by understanding what the base system prompt already covers
  • Choosing which Claude model to use for specific tasks based on their foundational instructions
  • Anticipating how future updates might change Claude's behavior
  • Learning prompt engineering patterns from Anthropic's own examples

Clone the repository and spend 10 minutes exploring—you'll gain insights that would take hours to extract manually from Anthropic's documentation.

Source: gentic.news · · author= · citation.json

AI-assisted reporting. Generated by gentic.news from multiple verified sources, fact-checked against the Living Graph of 4,300+ entities. Edited by Ala SMITH.

Following this story?

Get a weekly digest with AI predictions, trends, and analysis — free.

AI Analysis

Claude Code users should immediately clone Willison's repository and run `git log --oneline` on the prompt files for models they use regularly. This reveals the evolutionary path of Claude's "personality." Specifically, compare the system prompts for Claude 3.5 Sonnet versions you've used in production. Look for additions related to code quality, security, or reasoning patterns. These changes might explain subtle differences you've noticed in Claude Code's output over time. When Anthropic releases a new model version, check the git diff before the official release notes. The system prompt changes often reveal the engineering priorities—whether they're emphasizing security, creativity, or precision. Adjust your `CLAUDE.md` accordingly to complement (not duplicate) what's already in the system prompt.
Compare side-by-side
Anthropic vs GitHub
Enjoyed this article?
Share:

AI Toolslive

Five one-click lenses on this article. Cached for 24h.

Pick a tool above to generate an instant lens on this article.

Related Articles

From the lab

The framework underneath this story

Every article on this site sits on top of one engine and one framework — both built by the lab.

More in Open Source

View all