Claude Code Adds AI-Drafted Feedback Reports via /feedback
Claude CodeAnthropic shipped a new SendFeedback tool in Claude Code 2.1.247, giving Claude the ability to draft a feedback report on its own when something goes wrong in a session, a tool keeps failing, or the user points out a mistake. The draft surfaces through the /feedback command for the user to review before anything is sent, and the behavior can be turned off entirely with the new feedbackDrafts setting. The feature is meant to shorten the loop between hitting a problem in a live session and getting it in front of Anthropic, without requiring the user to leave their work to write up what happened.
Key Takeaways
- Claude can now self-initiate feedback drafts, triggered by a failing tool, a self-caught mistake, or a user calling out a problem, rather than requiring a user to write the report from scratch.
- Drafts route through
/feedbackfor review, so nothing is sent to Anthropic without the user explicitly approving the content first. - A new
feedbackDraftssetting provides an opt-out, letting users or organizations disable automatic drafting and keep the fully manual flow. - Independent testing flagged a real privacy risk almost immediately: a GitHub issue documents a draft that included private org names, repo URLs, and file paths before the user intervened.
- The feature targets a longstanding gap in developer tooling, where in-the-moment context about a bug is usually lost by the time a user gets around to reporting it manually.
- It shipped alongside a related UX nudge in the same release, a one-keystroke prompt on Bash permission dialogs pointing users toward auto mode, showing Anthropic's broader push toward reducing manual overhead in the CLI.
Sources & Mentions
1 external resource covering this update
Claude Drafts Its Own Feedback Reports
Claude Code 2.1.247 introduces SendFeedback, a built-in tool that lets Claude notice when something has gone wrong in a session and draft a feedback report about it without being asked. Previously, filing useful feedback meant a user had to stop what they were doing, reconstruct what happened, and write it up themselves, often losing the specific context (which tool failed, what the error looked like, what Claude tried) by the time they got around to it.
With the new tool, Claude can draft the report in the moment: when a tool keeps failing, when Claude catches its own mistake, or when the user calls out a problem directly. The draft is not sent automatically. It surfaces through the /feedback command, where the user reviews the content before deciding whether to send it. This keeps a human in the loop on anything that leaves the session, which matters given that feedback drafts can reference file paths, error output, or other session-specific detail.
Opt-Out Control
Organizations or individuals who would rather Claude never draft this kind of report on its own can turn the behavior off completely with the new feedbackDrafts setting, restoring the previous fully-manual /feedback and /bug flow.
Why It Matters
The gap between "something broke" and "Anthropic hears about it" has historically been one of the biggest sources of lost signal for a fast-moving CLI tool: users hit a bug, work around it, and move on without ever filing it. By having Claude itself notice and pre-draft the report at the moment of failure, Claude Code lowers the effort required to close that loop. Early independent testing has already surfaced a real edge case worth watching: a GitHub issue filed against the repository documents Claude Code drafting a public bug report that included real organization names, repository URLs, and file paths from a private codebase, information the user had to catch and redact before submission. That underscores why the review step in /feedback is not cosmetic: automated drafting still needs a human check before anything leaves the local environment.