Summary & Key Takeaways ​

• Simon Willison developed a method to extract and track the system prompts of various Claude models (Opus, Sonnet, Haiku).
• The process involves querying the Anthropic API, extracting the core system prompt from the tool_use_system_prompt field.
• He then commits these extracted prompts to a Git repository, creating a historical timeline of changes.
• This "git blame" approach allows developers to observe how LLM providers fine-tune model instructions over time.
• The article includes the Python script used for this extraction, enabling others to replicate the process.

Our Commentary ​

This is brilliant. The idea of using Git to track LLM system prompts is such a simple yet powerful concept. It addresses a real pain point: the lack of transparency around how these models are updated. We're constantly adapting our prompts and applications to new model versions, and having a historical record of the underlying system instructions could be a game-changer for debugging and understanding behavioral shifts. It's a fantastic example of community-driven transparency, and we hope more people adopt similar methods for other models. This kind of open-source tooling is exactly what the AI ecosystem needs.