I am associated with DnD Map Maker (https://dndmapmaker.com/); this is a disclosed project-related design question, not an independent security review. AI assisted with drafting this message. I am not claiming that the tool implements the privacy properties below. Layout and downloads are free, while AI generation and editing require login and credits. The question is how a small map-sharing workflow should separate facilitator-only information from the artifact handed to players. A hidden room name is not particularly sensitive in a global sense, but it is an understandable example of a disclosure boundary: the same picture may need a private working copy and a public, deliberately reduced copy. I think this is a useful miniature of a broader document-sharing problem. My proposed threat model is intentionally limited. Assume the recipient is allowed to inspect the downloaded file, including its text, filenames and metadata. Do not assume that a label is secret merely because it is invisible in the current viewport. Also distinguish a client-side annotation being hidden from that annotation never being included in the delivery artifact. A screenshot may reduce some kinds of embedded information, but it can still disclose a private label or a revealing filename; it is not a general privacy guarantee. A possible workflow is to maintain a separate player-facing source rather than repeatedly exporting the full facilitator source with a few visibility switches changed. Give both copies the same revision identifier, but separate their audience labels. Review the actual delivered artifact rather than only the editor preview. Retain the previous accepted player copy so that an experimental edit does not silently become the new default. Avoid putting real player names or private meeting details into prompts when fictional labels will do. I am interested in where people would enforce this boundary. Would you prefer a deliberately separate public document model, an export-time allowlist of permitted information, or both? What small check would catch the common failure of sending the right-looking file with the wrong underlying information? I have not run a comparative security test or measured these alternatives, so the examples are proposals rather than results. The general issue is minimizing what crosses an audience boundary, not making a product-specific security claim.