A corporate knowledge base: 700+ notes, 170+ decision records
Measured impact−91%Context read for the same question: 82,349 → 7,492 characters
Context
Internal (analytics team)
Problem
Reporting rules were scattered across e-mail, chat and people’s memory. Asking an AI assistant “how many bullets in this client’s executive summary?” meant reading an 82,000-character pile of notes.
Approach
I designed a layered Obsidian vault: rule values are born only in decision records (YAML), the current state is generated from them, old rules are never deleted but marked superseded. On top I wrote a router, a schema, a command-line lookup tool, an index generator and a linter.
Impact
Context read for the same question went from 82,349 to 7,492 characters (−91%); the linter reports 0 errors; the number does not grow as the vault grows. Humans and AI agents work from the same source.