Claude system prompts as a git timeline

The Catalyst: Why Does This Matter? Ever wondered what defines the "personality" and "capability boundaries" of chatbots like ChatGPT or Claude? The answer often lies in the "system prompt." Think of it as the initial instruction set given to an AI, dictating what it can and cannot do, and how it should speak. Historically, these prompts were a black box. Now, companies like Anthropic are partially disclosing them. However, the disclosure is often a single, monolithic Markdown page. Trying to understand how it evolves from this static document is like reading a book without a version history—painfully cumbersome. Simon Willison did something deceptively simple yet profoundly insightful: he took Anthropic's published Claude system prompt history, used Claude Code to break it into individual files organized by model, version, and timestamp, and placed it in a Git repository. This immediately brought the dry documentation to life. Deconstruction: What's the Core Method? The brilliance of this approach lies in engineering the documentation. Instead of manually comparing different Markdown versions, Willison let AI (Claude Code) handle it automatically: 1. Structuring: Splitting the single page containing all prompts for all models (like Opus, Sonnet) into a clear directory structure, e.g., /claude-opus/4.6.txt, /claude-opus/4.7.txt. 2. Versioning: Creating a Git commit for each prompt change, stamped with a fake but logically consistent timestamp. Thus, the entire evolution of the prompts becomes a clear Git timeline. What does this unlock? You suddenly have access to the developer's most familiar arsenal: use git log to see the history, git diff to pinpoint exactly which words changed between Opus 4.6 and 4.7, and git blame to see when a specific instruction first appeared. Willison himself used this tool to easily write a detailed analysis of the changes from Claude Opus 4.6 to 4.7. It's no longer a vague feeling that "the new version seems different," but a precise, character-level audit of changes. Trend Insights: What Larger Trends Does This Reveal? This project illuminates three deeper trends: First, "Configuration as Code" for AI Products. System prompts are becoming the "core configuration" of AI applications, their importance rivaling that of code or config files in traditional software. Their management, auditing, and iteration deserve the same rigor. Introducing the Git methodology is bringing software engineering best practices (version control, diffing, change tracing) into the layer that defines AI behavior. Second, The Engineering of AI Explainability and Debugging. When an AI behaves unexpectedly or not as intended, how do you troubleshoot? Comparing system prompt changes becomes an extremely efficient, low-cost entry point. This tool provides a concrete, actionable engineering paradigm for "AI behavior debugging." It tells us that understanding AI doesn't always require delving into complex neural networks; starting with the "instructions" it receives is often more direct. Third, Open-Source Intelligence (OSINT) Applied to AI. Using public information (a company's published prompt documents) combined with a clever toolchain, one can gain deep insights into a top AI company's product iteration thinking and security strategy adjustments. This offers the entire community a new methodology for researching, supervising, and understanding frontier AI models. Practical Value: How Does This Relate to Me? If you are an AI application developer or product manager, this method is directly applicable. When iterating on your own AI product, could you manage your system prompts similarly? This allows you to clearly see exactly what each adjustment changes, facilitating rollbacks and team collaboration. If you are an AI researcher or tech enthusiast, this tool and the thinking behind it are a treasure trove. You can use it to study the design philosophies of different companies' prompts, observe how safety rules are incrementally strengthened, or analyze what instruction-level optimizations correspond to model capability improvements. For instance, through a diff, you might discover that a new version quietly added a stricter instruction regarding handling sensitive information. Even if you're just a casual user, this case makes you realize: an AI's "personality" and "behavior" are meticulously designed and continuously adjusted. Next time you feel "Claude seems smarter" or "the response style changed," it might be the result of a system prompt update. The Counterintuitive Angle An angle that might be overlooked is that this project itself uses AI (Claude Code) to build tools for researching AI. Willison didn't manually write a script to parse the Markdown; he had Claude do the splitting and structuring. This creates an interesting recursion: using AI to understand and audit AI. It foreshadows that when developing AI-related tools in the future, the greatest leverage might come from cleverly utilizing existing AI capabilities themselves. In summary, Simon Willison's small project acts like a delicate key, opening a door for us to systematically understand the inner workings of AI products. It's not just a study about Claude; it offers the entire industry a new way to operationalize and engineer the "black box" of AI.