What they did
Ramm already had Claude Code maintaining a personal wiki of over a thousand Markdown pages, built from a folder of raw sources and read in Obsidian. Each time Claude ingested a source it logged what the source left unanswered, and that list had grown past a hundred lines with nothing ever closed. He built a small local web view, the Vault Cockpit, that turns each question into its own file with an owner, status and date. He types a short answer and picks what happens next: Claude works it in, follow up on a date, close, or record it as a decision. His words are saved as evidence in the raw folder first. He then tells Claude Code to work through the cockpit inbox; a skill has it read each answer, update the affected wiki pages with a citation and mark the item done. Claude also adds new questions through the tool's command line during ingests and proposes items to close as stale or duplicate.
How it went
On the Saturday he switched it on he answered close to 280 questions, about one a minute, while Claude worked the inbox a few minutes behind him. By evening more than 200 answers were in the wiki, and Claude had suggested close to 40 triage verdicts, about half of them stale. A few days later two items were still open and eleven were parked until a date. He found about one in five of the questions were contradictions between sources.
Worth knowing
The cockpit itself calls no model; nothing happens until someone pastes the prompt into Claude Code or a scheduled run picks it up. Each change runs as one transaction with a lint check and a local Git commit, and is rolled back if the check gets worse. Only the human can mark a decision valid or withdrawn. The tool is MIT-licensed on GitHub with a demo vault.