Claude Reflect: see what you delegate to AI, not just how often you use it
Anthropic's new Reflect feature turns chat history into a usage report. The interesting part is not screen time but the question of which decisions you still want to keep for yourself.
AI tools have rapidly increased usage volume, but the quality of that usage is hard to measure. Knowing how many messages I send each day does not tell me which work improved or which thinking habit weakened. Anthropic's Reflect feature, launched in beta on July 9, looks directly at that gap.
Reflect turns Claude history into a usage summary. It shows the topics, task types, and patterns in the last one, three, six, or twelve months. You can add break reminders and quiet hours on top of that view.
The real metric is delegation, not time
The strongest prompt in the feature is roughly: what is one thing you want to keep doing yourself, even if Claude could do it faster? That is more valuable than a screen-time chart. The main risk in AI use is not simply using it often; it is quietly delegating work that requires judgment.
For example, collecting research sources can be delegated while deciding which sources are trustworthy should stay with me. Draft variations can be generated while the final tone remains mine. A code change can be prepared while irreversible publishing or data operations require explicit approval.
A four-dimensional self-review
Anthropic ties the report to its 4D AI Fluency framework: Delegation, Description, Discernment, and Diligence. This is broader than 'write better prompts' because it treats deciding what not to delegate as part of AI skill.
A privacy note
Reflect works for users with Memory enabled. Anthropic says incognito chats, underlying files from connected tools, and conversations tied to health integrations are excluded. A usage summary can still be personal and sensitive, so it is worth checking what history is retained and who shares access to the account before enabling it.
This matches the conclusion in my own note on what AI changed in my workflow: as speed increases, the boundaries around decisions need to become more deliberate.
I would use Reflect as a monthly retrospective, not a performance score. Three questions are enough: where did AI genuinely save time, which outputs did I accept too easily, and which decision will I keep for myself next month? The feature's value is not answering those questions but making them recurring.