Initialize a personal knowledge base using the LLM Wiki pattern: the LLM reads raw sources, builds a persistent wiki of summaries/entities/concepts/analyses, and keeps it current. You curate sources; the LLM does the bookkeeping.
Core principle: The wiki is a compounding artifact — cross-references, contradictions, and synthesis are already done. No embedding-based RAG needed.
Quick Reference
| Path | Owner | Purpose |
|------|-------|---------|
| raw/ | Human | Immutable source docs |
| wiki/ | LLM | Generated pages |
| AGENTS.md | Human + LLM | Schema + runbook |
| wiki/index.md | LLM | Content catalog |
| wiki/log.md | LLM | Activity log |
| wiki/overview.md | LLM | Evolving synthesis |
Initialization
When asked to initialize a wiki, create this structure:
.
├── AGENTS.md # Schema + agent runbook (co-evolved)
├── raw/ # Immutable sources (human-only writes)
│ ├── specs/ # Technical docs, architecture
│ ├── papers/ # Research papers, whitepapers
│ ├── vendor/ # Vendor docs, product pages
│ └── assets/ # Downloaded images/diagrams
└── wiki/ # LLM-generated (LLM writes)
├── index.md # Master content index
├── log.md # Append-only activity log
├── overview.md # High-level synthesis
├── summaries/ # One file per ingested source
├── entities/ # Systems, vendors, standards, tools
├── concepts/ # Technical concepts, patterns, principles
└── analyses/ # Comparisons, syntheses, deep dives
AGENTS.md
Create with: directory structure + ownership rules, page templates, workflows, conventions, domain considerations. This file turns the LLM from a chatbot into a disciplined wiki maintainer.
wiki/index.md
Content catalog organized by category. Each entry: link, one-line description, optional metadata.
wiki/log.md
Append-only: ## [{date}] {action} | {title}. Actions: ingest, query, lint, update, create.
Page Templates
| Type | For | Key fields | |------|-----|------------| | Entity | Systems, vendors, standards, tools | type, status, overview, key facts, relevance, related, sources | | Concept | Patterns, principles, techniques | category, definition, why it matters, how it works, trade-offs, examples, related, sources | | Summary | One per ingested source | source path, type, ingested date, TL;DR, key takeaways, detailed notes, questions raised, connections | | Analysis | Comparisons, syntheses, decision matrices | type, dates, question, methodology, findings, recommendations, confidence level, sources |
Workflows
| Workflow | Steps | |----------|-------| | Ingest | Read source → discuss takeaways → create summary → update/create entity + concept pages → cross-reference → update index + overview → log | | Query | Read index.md → read relevant wiki pages (not raw) → synthesize with citations → consider filing as analysis → log | | Lint | Check orphans, stale claims, contradictions, missing pages, missing cross-refs, data gaps |
Conventions
- File naming: lowercase, hyphen-separated (
foobar-standard.md) - Links: relative markdown —
[FOOBAR](../entities/foobar-standard.md) - Dates: ISO 8601 (
2026-04-14) - Citations: every claim traces to a source summary
Reference
Full pattern: ./references/LLM-WIKI.md.