Rolling Out to Your Team
This guide covers how to move from a passive wiki to an active, self-updating knowledge base where AI agents capture knowledge during normal work.
The core idea
Instead of asking people to stop working and write documentation, each person connects their AI tool to Lore DB via MCP. The AI inherits their permissions, can read existing docs to give better answers, and can write back new insights as they emerge. Documentation becomes a byproduct of work, not a separate task.
How knowledge capture works
| Mode | How it works |
|---|---|
| Passive retrieval | Someone asks a question. The AI searches Lore DB and answers with real documentation. |
| Prompted capture | The AI detects new knowledge in conversation and offers to save it. “That fix isn’t documented yet. Want me to add it?” |
| Explicit authoring | The user tells the AI to create or update a document directly. |
Security: why individual connections matter
Each person’s MCP connection authenticates as that specific user via Google OAuth. The AI inherits their exact library permissions. This means:
- A Viewer’s AI can search and read, but write requests are rejected by the server
- An Editor’s AI can only modify libraries they have Editor access to
- All writes are attributed to the individual — full audit trail via the
updated_byfield - Revoking access is instant. Remove someone from a library and their AI loses access immediately
The alternative — a shared bot with a master API key — is a security risk. One prompt injection or misconfigured command could modify the entire knowledge base with no attribution. Individual connections eliminate this class of risk entirely.
Recommended library structure
| Library | Access | Contents |
|---|---|---|
| Core Playbooks | All staff: Viewer. Knowledge champions: Editor. | Deployment runbooks, incident response, onboarding checklists |
| Team: [Name] | Team members: Editor. Other teams: Viewer. | Architecture decisions, service guides, tribal knowledge |
| Pending Insights | Reviewed weekly. AI-captured content lands here first. | Incident findings, draft procedures, observations |
| Personal | Auto-created per user. Private scratch space. | Meeting notes, draft proposals, quick references |
The flow: knowledge starts in Pending Insights or a personal library → gets reviewed by a knowledge champion → gets promoted to the relevant Team or Core library.
Rollout phases
Pick one team. Set up their library + Pending Insights. Grant Editor to 3–5 people. Connect their AI tools. Knowledge champion does the first weekly review at end of week 1.
Add 2–3 more teams. Create Core Playbooks library. Enable cross-team Viewer access. Refine AI agent instructions based on pilot feedback.
All teams onboarded. “Connect to Lore DB” becomes a day-1 onboarding task. Monthly metrics review begins.
See the Quality Control workflow for how to run the weekly review effectively.