Zero to operator · 8 chapters
Build Your Own LLM Wiki
Turn scattered chats and documents into an inspectable knowledge system that compounds.
A source-first field guide to durable Markdown, explicit schemas, safe ingestion, cited retrieval, knowledge graphs, maintenance, and multi-agent access. You begin with folders and files, then progressively build the machinery that makes the wiki useful and trustworthy.
Begin chapter oneDeliberate progression
The curriculum
Foundation chapters give you the map. Builder chapters teach you to make controlled changes. Operator chapters teach permissions, recovery, quality, and proof.
- 01FoundationKnowledge That Outlives the ChatUnderstand what an LLM Wiki is, what it is not, and why readable files are the foundation of trustworthy AI memory.26 min lesson + lab
- 02FoundationDesign a Vault Humans Can ReadCreate a file and schema system that stays understandable before any automation exists.32 min lesson + lab
- 03FoundationCapture Sources Without Losing ProofPreserve originals, context, permissions, and the line between quotation and inference.36 min lesson + lab
- 04BuilderBuild the Ingestion PipelineTurn source material into proposed linked pages through a pipeline you can test one stage at a time.45 min lesson + lab
- 05BuilderRetrieval That Can Show Its WorkRetrieve evidence first, synthesize second, and report uncertainty instead of hiding it.49 min lesson + lab
- 06BuilderLink, Lint, and ConsolidateKeep the wiki coherent by treating links, duplicates, contradictions, and staleness as maintainable state.44 min lesson + lab
- 07OperatorConnect Agents Without Surrendering ControlLet Claude, Codex, and other tools share the wiki through narrow interfaces, policy, and reviewable writes.47 min lesson + lab
- 08OperatorOperate, Evaluate, and RecoverMake the wiki dependable with evaluations, quality metrics, backups, migrations, incident procedures, and visible trust.53 min lesson + lab
Capstone
Ship a Source-Backed Personal Knowledge System
Build a local-first LLM Wiki from at least twenty real sources. A second person or agent must be able to ingest material, navigate the vault, ask questions, inspect citations, identify gaps, review proposed changes, and restore the system without relying on your memory.
What you must build
- ◆Readable map, root instructions, versioned schemas, stable IDs, and canonical-page rules.
- ◆Preserved originals with envelopes, access scopes, checksums, and derivation records.
- ◆Idempotent text ingestion with structured proposals, identity resolution, dry run, and receipt.
- ◆Lexical retrieval plus optional semantic search with visible fallback.
- ◆Answer contract with citations, path, confidence, freshness, gaps, and verification step.
- ◆Graph and linter covering links, orphans, duplicates, stale pages, sources, and contradictions.
- ◆Agent access policy with least privilege, revisions, proposals, approvals, and receipts.
- ◆Ten-question evaluation, backup, clean restore, and incident runbook.
How you prove it
- Recording or transcript from source capture through cited answer.
- An unanswerable query returning an honest useful gap.
- Repeated ingest proving no duplicate creation.
- Rejected unauthorized or stale write.
- Before and after maintenance report with reviewed diff.
- Clean-room restore ending in a passing query.
