llm-wiki-pattern

Source: LLM Wiki — a pattern for building personal knowledge bases using LLMs

The seed essay for this whole repo. Argues that mainstream LLM-document workflows (RAG, ChatGPT file uploads, NotebookLM) rediscover knowledge from scratch on every query, never accumulating anything. The alternative: have the LLM incrementally maintain a persistent wiki of interlinked markdown files between you and the raw sources, so knowledge compiles once and stays current.

Key takeaways

• The wiki is a persistent, compounding artifact, not just an index. See compounding-knowledge.
• Three layers: raw sources (immutable) / LLM-maintained wiki / schema file (CLAUDE.md). See three-layer-architecture.
• Three operations: ingest (raw → wiki), query (ask → answer, optionally filed back), lint (health check). See ingest-query-lint.
• Indexing: index.md is content-oriented; log.md is chronological with grep-able prefixes.
• Maintenance is the bottleneck humans abandon; LLMs don’t, which is why this can finally work. See llm-wiki-pattern.
• Spiritual lineage: Vannevar Bush’s memex (1945) — Bush couldn’t solve who-does-the-maintenance; the LLM solves that.
• Tooling: Obsidian as the read-side UI; the LLM as the write-side. qmd for search at scale. Web Clipper for ingest.

Quotes worth keeping

“Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.”

“Humans abandon wikis because the maintenance burden grows faster than the value. LLMs don’t get bored.”

How this seeded the repo

The CLAUDE.md schema in this vault is a concrete instantiation of the pattern: directory layout, frontmatter conventions, ingest/query/lint workflows, wikilink discipline.

Open questions raised

• At what scale does the index file stop being enough and qmd (or similar) become required?
• How do we handle the LLM’s read-then-image step for clipped articles with inline images cleanly?
• What’s the right cadence for lint — weekly, or driven by some signal?