The provided text outlines the architecture and management of "system prompts" employed within Claude’s web interface and mobile application. These system prompts serve as a foundational instruction layer, dynamically injected at the initiation of every conversation to provide the model with essential context, such as the current date, and to mandate behavioral constraints, such as the requirement for Markdown-formatted code blocks.

The methodology involves a continuous refinement process where these system prompts are updated periodically by the developers to enhance the model's response quality, accuracy, and adherence to user-defined formatting preferences. A critical distinction is made regarding the deployment of these prompts: they are exclusive to the web and mobile interfaces and are not applied to the Claude API, ensuring that API users maintain full control over their own system-level instructions.

Regarding versioning and model management, the document categorizes Claude models into two distinct architectural approaches. For models released from Claude 4.6 onward, the system utilizes a "fixed snapshot" methodology. In this framework, each model ID represents a static, immutable version of the model, resulting in a single configuration entry for each ID. Conversely, earlier versions and specific iterative deployments allow for multiple system prompt entries per model ID, reflecting a more fluid updating schedule. The text concludes by listing various model iterations (e.g., Fable 5.1, Opus 5, Sonnet 3.7) to illustrate the evolution of these specific architectural snapshots over time.