Claude Code's New Feature: It Drafts Its Own Feedback Report When Something Goes Wrong, and You Decide Whether to Send It

How is this automatic drafting different from the existing /feedback and /bug commands?

The existing /feedback and /bug commands were a passive flow — you only open them once you actively think to report something. You have to first recognize there's a problem, remember the command, and organize what you want to say yourself. The new automatic drafting instead hands off the judgment of "noticing a problem" to Claude itself, which proactively triggers it under specific conditions: a tool keeps failing, a task doesn't get completed, or either side points out a mistake. The draft gets written first and placed in a queue, waiting for you to decide.

The two aren't a replacement relationship — they coexist. If you'd rather write it entirely yourself, pressing w still opens the standard feedback window, or you can go directly through /feedback or /bug with text content — that path is completely unchanged. The only difference is there's now an additional path where Claude proactively drafts it for you first, so you don't have to describe what just happened from scratch every time.

Why is it specifically emphasized that "nothing goes to Anthropic without you actively sending it"? What design consideration does this reflect?

The nature of automatic drafting is that Claude organizes "what just happened" into text before you may have even noticed there was a problem. If that draft got sent automatically before you reviewed it, Claude would essentially be deciding on its own that "this is worth reporting," with no room for you to weigh in at all — unacceptable to many developers, especially since a draft might contain sensitive content like project context or code snippets.

Splitting "drafting" and "sending" into two clearly separate actions, with your active confirmation required in between, keeps the decision of whether Anthropic gets to see this entirely in your hands. This also explains why organizations with Zero Data Retention enabled don't get this feature at all — the /feedback command itself is unavailable to them, because the core principle of that kind of organization is not generating, or retaining, any data that could potentially be sent out. Even though automatic drafting technically doesn't auto-send anything, the mere act of generating and storing a draft locally already conflicts with the spirit of Zero Data Retention.

If I'm using Claude Code day to day, what will I actually see, and what do I need to do?

The first time it triggers, you'll see a card above the input box showing the draft's title — that's your first decision point: press 1 to view the details, press 2 twice to send it as-is, or press 0 to dismiss the card (the draft doesn't disappear, it just stops showing the reminder card). If you dismiss the card, Claude Code then asks whether you'd like to turn off automatic drafting entirely; decline twice in a row and it stops asking.

If you want to systematically go through everything that's accumulated — cards you dismissed, or ones you never even saw — typing /feedback with no text opens the full queue, where you can edit each draft's title, category, and content one by one, decide whether to include the transcript, and send or delete it. Note that the queue itself has a cap: at most 10 drafts total, and once an 11th gets created, the oldest one is automatically cleared. Drafts left unprocessed also expire automatically after 30 days — meaning this feature isn't designed to work as a permanent to-do list, but rather as "handle it while you still remember what happened."

Could this feature send sensitive project content out without me knowing? How do I make sure it's safe?

By design, the answer is no — the whole process has two layers of protection. First, drafts only ever exist locally at ~/.claude/feedback/drafts/; nothing leaves your computer until you actively press send. Second, even once you decide to send, whether to include the transcript is a separate, independently toggleable option — it's not bundled in and force-sent.

What's actually worth paying attention to in practice is what the draft actually contains — since it's automatically organized by Claude, it may include code snippets, file paths, or project context from that moment. It's worth actually opening and reviewing it before sending, rather than pressing 2 twice in a row without looking. If your organization is particularly sensitive about this kind of data leaving the machine, you can disable automatic drafting through settings, or evaluate whether Zero Data Retention should be enabled — once it is, both the /feedback command and this entire drafting tool become unavailable outright, removing the concern at the source.