Anthropic announced on August 27, 2026, that Claude Code can now automatically draft bug reports when errors occur, with all drafts stored on your machine until you explicitly approve sending them. The feature, called Claude-drafted feedback, has sparked immediate questions about data privacy and token usage among developers on rate-limited plans.

What Exactly Happens When Claude Drafts a Bug Report?

When Claude Code encounters a failure, notices its own mistake, or detects user frustration, it writes a structured bug report and saves it to a local directory on your machine at ~/.claude/feedback/drafts/. Nothing transmits to Anthropic's servers automatically. Instead, you see a card above your prompt with three options: press 1 to review the draft, 2 to send it, or 0 to dismiss it entirely. Only when you press 2 does the report leave your machine.

The feature requires Claude Code version 2.1.238 or later, which shipped on August 20, 2026, a week before the public announcement. Interestingly, the tool was quietly present in releases before the X post made it visible, discovered through Anthropic's tools reference documentation rather than release notes.

Does This Feature Drain Your Token Budget or Rate Limits?

The short answer is no, not meaningfully. The SendFeedback tool operates as a normal in-session tool call, meaning you pay only for the tool definition itself plus a few hundred words of draft text. The system caps drafts at three per session by default, and Anthropic can adjust this limit server-side without requiring a software update. Critically, sending a draft runs zero additional inference, so it consumes no extra tokens beyond what you've already used in your session.

For developers on usage-capped plans, this distinction matters. You're not triggering a separate model call to generate the report; Claude writes it as part of your existing conversation, and transmission costs nothing extra.

What Data Gets Included in the Report?

Each draft follows a rigid, structured format designed to match professional bug-report templates. The report includes the following fields:

• Title: A concise name for the issue, such as "Sandbox image pull fails behind proxy."
• What Happened: Observed behavior versus expected behavior, with exact error text if it's short enough to include.
• What the User Said: Your own words, quoted directly from the conversation, or a note that the model observed the issue without user comment.
• Repro: The minimal steps or conditions needed to reproduce the problem.
• Evidence: Chaseable identifiers like request IDs, timestamps, file paths, and version numbers, omitted only if none exist.
• Cause: Optional field, included only if Claude verified a root cause during the session.

Behind the scenes, each draft also carries classification fields that never appear on the card you see. These include a report type (bug, idea, or missing_capability), an area name (such as "hooks config" or "file editing"), and two enums that reveal what Anthropic is actually measuring: failure_mode and other diagnostic categories.

One critical detail: the transcript of your entire conversation is omitted from all card-sent reports by default. You can optionally enable "Send transcript: yes" on the review screen if you want to include it, but this is not the default behavior.

How to Control or Disable Claude-Drafted Feedback

• Turn It Off Completely: Run /config, navigate to Claude-drafted feedback, and select off. This writes "feedbackDrafts": "off" to your configuration file.
• Disable for One Session: Set the environment variable CLAUDE_CODE_SEND_FEEDBACK=0 before launching Claude Code to skip the feature for that session only.
• Keep Drafting Without Cards: Use the "quiet" setting, which continues to generate drafts but suppresses the review cards, leaving you to find them manually in ~/.claude/feedback/drafts/.
• Non-Interactive Environments: The SendFeedback tool is completely omitted from command-line runs using the -p flag, Agent SDK sessions, Claude Code on the web, and all Bedrock, Vertex, or Foundry deployments, so no drafts are created in those contexts.

Why Is Anthropic Building This Feature Now?

The tool's own instructions reveal a telling list of self-reporting cases that Claude should flag: giving a confident answer then retracting it, stopping short when it could have finished, declining a reasonable request, spawning too many subagents, using the wrong tone, asking excessive clarifying questions, or expanding scope beyond what was asked. This catalogue reads like a direct response to months of complaints on social media about Claude Opus 5 feeling worse to work with, including over-clarification, premature stopping, and scope creep.

Anthropic did not build a traditional bug tracker here. It built a telemetry channel for model-behavior regressions, aimed squarely at failure modes people have complained about but rarely bothered to file formally. The feature includes two guardrails to prevent spam: the three-draft-per-session cap and an explicit instruction to draft only at natural moments and never re-draft the same issue twice in one session.

There is also a deliberate silence rule. The tool renders no UI and does not interrupt the conversation, so Claude will never announce that it filed something. The card above your prompt is the only signal you receive.

What Does This Mean for Enterprise AI Deployments?

The Claude-drafted feedback feature reflects a broader shift in how Anthropic approaches enterprise AI. On the same day, Anthropic and Salesforce announced Claudeforce, an integration that connects Claude's reasoning capabilities with Salesforce's customer data, business logic, and workflows. Salesforce in Claude launches with 37 prebuilt sales skills covering meeting preparation, deal health reviews, and pipeline reviews.

The distinction between these two developments is important. Claude-drafted feedback is a feedback loop that helps Anthropic understand where its model is failing. Claudeforce, by contrast, is a controlled execution layer that allows Claude to take actions within existing business systems while respecting permissions and approval processes. Both reflect a maturation of AI from a writing assistant sitting beside applications toward a system that integrates deeply with enterprise operations.

For organizations evaluating Claude Code or other AI agents, the feedback feature demonstrates Anthropic's commitment to transparency and user control. Drafts stay on your machine until you act, nothing drains your token budget, and you can disable the feature entirely if you prefer. The structured format and local-first design suggest Anthropic is thinking carefully about how to gather signal from real-world failures without creating privacy or cost concerns for developers.