Anthropic has officially launched the test version of the Reflect feature, turning Claude into a "mirror" to help users review and reflect on their AI usage habits. This feature is available to free, Pro, and Max plan users who have enabled the memory function, with access through the settings page on the web and desktop applications.

Users can choose to review chat activities from the past 1 month, 3 months, 6 months, or 12 months. The system will automatically generate a personalized usage summary. The report covers key topic distribution, frequent usage patterns, common task types, as well as the user's most active time periods and corresponding work content. Anthropic also revealed that future updates will include a visual timeline of usage duration, support for setting silent periods, and reminders about usage duration.

image.png

Four Dimensions of "Checkup," From Using AI to Mastering AI

The core design of Reflect is to comprehensively evaluate users' AI collaboration habits across four dimensions. These four dimensions are task delegation, goal description, result identification, and responsible caution. Each dimension provides specific collaboration examples and targeted improvement suggestions. This is equivalent to providing each user with an "AI usage health check report," telling you which tasks were assigned correctly, which instructions were unclear, and which results should be carefully reviewed before being trusted.

Anthropic's intention is clear: improving AI capabilities is not just about the model itself; users' own "questioning literacy" also determines the final outcome. While most people focus on what AI can do, Anthropic chooses to look inward and examine how well users are doing, which is a rare "internal perspective" in today's AI industry.

Clear Privacy Boundaries, Health Data Completely Excluded

In terms of privacy protection, Anthropic has clearly defined boundaries. Reflect will not read incognito chat content, nor will it extract underlying files from connected tools. For example, when a user asks Claude to summarize their inbox, the report may mention the behavior pattern of "users frequently summarize emails," but it will never include the original email content. Additionally, all dialogues related to health integration tools are completely excluded from the analysis scope.