Claude Code Introduces Auto-Draft Feedback Reports: Streamlining User Input for Anthropic

Anthropic’s Claude Code, the AI-powered coding assistant, has rolled out a new feature to simplify feedback reporting: auto-drafting feedback reports directly from user interactions. This tool aims to make sharing issues, errors, or suggestions more seamless for users.

The auto-draft function triggers in specific cases: repeated tool/command failures, unfulfilled user requests, self-detected or user-pointed errors, or explicit feedback requests. To access this feature, users need Claude Code v2.1.238 or later.

Privacy is prioritized here—all drafts are stored locally at ~/.claude/feedback/drafts/ and never sent to Anthropic until the user approves submission. The system retains up to 10 drafts; the oldest is deleted when the 11th is created.

When a draft is ready, users see a card above the input box with the draft title. They can press 1 to view it, double-press 2 to send as-is, or 0 to ignore. Ignored drafts stay in the queue, and after two ignores, the tool stops asking about disabling the feature.

Users can manage drafts via the /feedback command, which opens a queue of unprocessed drafts across sessions. From here, they can edit content, toggle transcript inclusion (default yes if session records exist), send, delete, or keep drafts. Manual feedback is available via /feedback with text or the /bug command.

The feature is unavailable in non-interactive modes, cloud sessions (no local storage), or platforms like Amazon Bedrock, Google Cloud Agent Platform, or Microsoft Foundry. It’s also disabled for Zero Data Retention (ZDR) organizations or if specific environment variables (e.g., CLAUDE_CODE_SEND_FEEDBACK=0) are set.

Sources

• Machine Heart Editorial Department, https://c.m.163.com/news/a/L5BBFI2E0511AQHO.html?from=subscribe
• Anthropic Official Documentation, https://code.claude.com/docs/en/tools-reference#sendfeedback-tool-behavior