Anthropic is building a personal scorecard into Claude that grades how effectively you interact with it. References to the feature have been found inside Claude's settings, where users will be able to open a dedicated screen and ask Claude to analyze their own conversation history.
The score covers 11 specific behaviors. Things like: do you tell Claude what you are trying to accomplish before asking for help? Do you specify the format you want? Do you push back when the reasoning seems off? Do you bring in context Claude could not otherwise know? Each behavior gets marked as demonstrated, partial, or not observed, and the final score comes out as a fraction, for example 7.5 out of 11, with notes on where to focus next.
The 11 behaviors come from real data. In February 2026, Anthropic published a study of 9,830 anonymized Claude conversations and found that the single strongest predictor of good AI use was iteration: people who refine and push back on outputs instead of accepting the first answer showed roughly double the number of good behaviors in the same conversation. The study also found a trap: when Claude produces something that looks polished, like a document or a piece of code, users become significantly less likely to check facts or question the reasoning. The scorecard is designed to flag exactly that.
What makes this unusual is that most AI tools are judged on how good the model is. Anthropic is instead trying to surface how good the user is. Only about 30 percent of users in the February study told Claude how to behave during a conversation, meaning most people never set the terms of the interaction at all. The scorecard would show that gap directly.
The feature fits into a broader set of moves Anthropic has been making on the education front. The Anthropic Academy already offers free courses built with academic researchers from University College Cork and Ringling College of Art and Design. There are separate tracks for small businesses, educators, students, and nonprofits, with a free course co-developed with PayPal specifically for small business owners. The scorecard would bring the same logic inside the product itself.
No launch date has been confirmed, and it remains unclear whether the feature will roll out to all users or start with paid and enterprise tiers. Either way, if it ships, it would be the first time a major AI lab built a feedback loop that grades the person asking, not the model answering.
For anyone using Claude at work, the underlying research is worth taking seriously regardless of the product feature. Checking facts, pushing back on logic, and iterating instead of accepting the first output are not habits most people develop on their own. Knowing your own score on those behaviors, based on actual past conversations, is a faster way to find out where time is being left on the table.