Response drafts let you preview Dosu's auto-generated responses before they're posted publicly.

Why Use Drafts#

Drafts give you a human-in-the-loop workflow where Dosu generates initial responses that you review and edit before posting. This is useful for teams that want to use Dosu as a research assistant while maintaining control over what gets published.

Drafts also help you build confidence in Dosu's responses before enabling fully automatic replies.

How Drafts Work#

When a deployment is set to Auto Draft mode, Dosu automatically generates responses to new issues and discussions. Instead of posting them immediately, it saves them as drafts for your review.

Drafts are only generated when certain conditions are met: the thread must be open, not already answered by Dosu, and not from a bot. Old or stale threads are also skipped unless someone explicitly mentions @dosu.

Important: When someone mentions @dosu directly, Dosu always posts a response immediately, even in Auto Draft mode. Drafts only apply to auto-triggered responses where no one mentioned Dosu.

Enabling Drafts#

To enable response drafts, set the reply mode to Auto Draft in your deployment settings. You can configure this separately for issues, discussions, and Slack channels.

For configuration details, see GitHub Configuration or Slack Configuration.

Reviewing Drafts#

Drafts appear on the Review page, where you can publish them as-is, edit them before publishing, or delete them if no response is needed.

Confidence Score Badges#

Preview messages on the Review page display a confidence score badge alongside the "preview" badge. This badge helps you quickly evaluate whether a draft response should be published or needs editing.

The confidence score reflects how confident the system is in the response quality based on the research conducted. The scores are categorized as:

• High confidence: The agent found relevant information during research
• Medium confidence: The agent found some related information during research; review is recommended before publishing
• Low confidence: The agent didn't find sufficient information during research; low confidence responses are never auto-published and are always held for review