Anthropic Publishes the System Prompts Running Behind Claude's Consumer Apps

Anthropic maintains a public release-notes page cataloging the system prompts that precede every conversation on claude.ai and its mobile apps. These prompts inject runtime context Claude wouldn’t otherwise have—most notably the current date—and steer default behavior, such as always returning code in Markdown. The page frames them as living documents that Anthropic revises periodically to improve responses.

The entry doubles as a changelog: it lists prompts per model, from older releases like Opus 3 and the Sonnet 3.5/3.7 line up through the current Opus 5, Fable 5, and Opus 4.8 generation, and bolds the diffs between dated versions so readers can track what changed. A key caveat is that these prompts govern only Anthropic’s first-party surfaces—the consumer web and mobile clients—and are not applied to the Claude API, where developers supply their own. From the Claude 4.6 generation onward, each model ID maps to a single fixed snapshot rather than an evolving series of dated updates.

For a technical audience, the significance is transparency. System prompts materially shape how a model behaves yet are usually invisible to end users, so documenting them—and their revision history—lets developers and researchers see exactly what instructions and context sit in front of user input, and distinguish consumer-app behavior from what they’ll actually get when building against the API.