What's this about?
When something fails, when Claude notices its own mistake, or when you tell it something went wrong, it now writes up the feedback report itself via a SendFeedback tool. You review, modify and approve before it sends. The engineer who built it said the reports genuinely help them find problems.
What exactly is the point of this?
Bug reports have a collection problem: the person best placed to describe the failure is the one least inclined to stop and write it up. Nearly all useful signal is lost to that friction. Having the tool draft the report — while the context is still live and complete — attacks the friction rather than asking users to care more. The approval step is what keeps it from silently exfiltrating your session.
What's the community saying?
2,091 likes on the announcement, 774 on the engineer's thread
- Approval for the review-before-send design, since the alternative would mean automatically uploading whatever was on screen.
- Interest in the self-noticing case — the model filing a report when it catches its own mistake, without being told.
- It fits the week's pattern of Anthropic shipping small friction fixes and publishing who built them.