Stack: Claude (Pro, web or desktop) + the monday.com MCP connector
Your board is not a record of the work. It is a record of the work somebody remembered to type in, in the fields they bothered to fill.
The gap between those two things is hidden work. It comes in three classes:
Class 1. On the board, invisible to every view. No owner, so it appears on nobody's My Work. No due date, so it never lands on a timeline or trips an overdue filter. Blank status, so the battery widget silently drops it from the denominator. Parked in a group nobody scrolls to.
Class 2. On the board, described incorrectly. Status reads "Working on it." Nothing has moved in six weeks. Or status reads "Done" while the last update says the client never signed off.
Class 3. Not on the board at all. A commitment made in an update thread. A four-line checklist buried in an item description. A blocker mentioned in a reply that never became an item.
Monday's own filters find Class 1. They cannot find Class 2 or Class 3, because a filter needs a predicate, and "the status column disagrees with what Amara wrote on the 14th" is not a predicate. That is the specific job you are handing to a language model.
Claude has no institutional memory of your project. It will produce confident findings that are wrong, and the wrong ones read exactly like the right ones. The single rule that makes this method safe is in Phase 3: every finding must cite an item ID and the exact phrase that triggered it, so you can check it in ten seconds.
This is a diagnostic, not an audit. On a large board you will be sampling, and a sample tells you what kind of problem you have, not how much of it there is.
Enable the monday.com connector in Claude, then:
Call
get_user_context. Tell me who I am authenticated as, my account plan, and my most-visited boards.
Resist the account-wide sweep on the first pass. Pick the board where a surprise would cost you most, usually the one that feeds a status report you have to defend.