What they did
Dan LeFebvre, who makes Jira training courses, was updating one 21-lesson section of a course and wanted a single board showing where every lesson stood. He connected Claude to a sandbox Jira Cloud site through Atlassian's official MCP server and handed it a screenshot of the outline from his notes app, a screenshot of his folder of finished videos, and a short note on version names. He asked it to build a tracker. Claude first came back with a plan: one epic, one task per lesson, which lessons looked finished based on the folder, and a rule for mapping two lesson numbering schemes that were off by one. He corrected the plan several times (a new space, no subtasks, his own status pipeline, and that unfinished lessons were already in production, not to do) before letting it create anything.
How it went
Claude could not create the Jira space or set up the statuses, so LeFebvre did that by hand in a few minutes. Once told the space was ready, Claude checked that the space and all four statuses existed, then made one epic and 21 tasks in roughly two minutes, about 45 calls in all, putting 8 lessons in Done and 13 in Production. Spot checks showed the right lesson numbers, statuses and parent epic on every card. The Claude app asked for approval before the first create, the first status change and the first move to Done. His sandbox showed zero Rovo credits used afterwards.
Worth knowing
Atlassian's MCP server can create, edit, comment on and move work items, but cannot create spaces or edit workflows, and it only works with Jira Cloud, not Data Center. There is no bulk create, so each item is a separate call. One of his corrections alone would have put 13 lessons in the wrong column, so read the plan before approving it. His Claude already had context on the course from earlier chats.