Knowledge System
SSOT and handbook-first culture.
PART 1: KNOWLEDGE SYSTEM — Single Source of Truth
Authoritative Resources
| Resource | Why It Matters |
|---|---|
| GitLab Handbook | Gold standard for "handbook-first" culture. Every single policy, process, and value documented publicly. All-remote operations model. |
| The Knowledge Manager's Handbook (Nick Milton) | Professional KM frameworks — the theory behind why SSOT matters |
| Notion's Company Wiki Playbook | Practical templates for early-stage knowledge bases |
Core Framework
The Handbook-First Principle (from GitLab): "If it's not in the handbook, it doesn't exist." Every discussion that produces a decision must result in a handbook update. This replaces tribal knowledge with institutional knowledge.
Key SOPs:
- Documentation-first workflow: Proposals written before implemented. Decisions documented after made.
- Template library: SOP templates per department (Standard Operating Procedures)
- Repository structure: /handbook → /engineering → /product → /sales → /finance → /legal → /people
- Review cadence: Quarterly stale-doc audit. AI agent flags outdated content.
- Access control: Public by default, sensitive docs locked. All AI agents read; designated agents write.
AI Integration
- AI agent that answers policy questions from the handbook (internal chatbot)
- AI that flags outdated docs based on tool data mismatches
- AI that generates first draft of SOPs from meeting notes and decisions
Three Knowledge Surfaces
Do not force every kind of knowledge into one tool. The company needs three deliberately connected surfaces:
| Surface | Canonical content | Consumer |
|---|---|---|
| Human handbook | Durable company policy, strategy, department guidance, and decisions | Humans |
| Operational knowledge base | Runbooks, architecture, procedures, references, and searchable implementation detail | Humans and agents |
| Agent memory | User preferences, temporal context, and working decisions needed for agent behavior | Agents |
The rule is one canonical home per fact. Cross-link between surfaces; do not copy the same content into all of them and hope they stay synchronized.
Evidence and Provenance
Every imported claim should carry enough provenance to answer: where did it come from, when was it checked, and is it fact, interpretation, or speculation? Mark claims as verified, partially verified, unverified, or opinion. Unverified claims can remain research notes, but they must not become policy, architecture, or customer-facing guidance without validation.
Curation and Publication Gates
- Curate, do not dump: preserve the source, extract durable concepts, file by subject, synthesize clearly, cross-link related pages, and record limitations. High-value material should not become a link-only entry.
- Orient before ingestion: read the schema, routing rules, current index/sidebar, activity log, and existing coverage before creating or changing a page.
- Navigation is data integrity: every page needs an index/sidebar entry, every entry must resolve to an existing canonical route, and orphan or broken pages must be reported.
- Publish through review: durable knowledge changes should pass through the normal branch, content/navigation checks, commit, and human review boundary rather than being silently edited into production.
- Separate failure classes: report content/MDX errors, navigation errors, resource limits, authentication/approval failures, and provider/network failures separately.
A page that exists on disk but cannot be found, trusted, or reviewed is not operationally complete.
Last updated on