Knowledge Wiki: integrate LLM Wiki pattern (Karpathy) into structured-context

[mindsocket]

Context

Karpathy's LLM Wiki gist (Apr 2026, ~1.2k stars) describes a pattern for building personal knowledge bases using LLMs — incrementally maintained, structured markdown wikis sitting between raw sources and queries. This aligns closely with what structured-context already supports (structured markdown spaces, schema validation, link integrity, agent skills) and extends it into a new domain: loose knowledge management alongside the stricter OST/product framework use case.

What this covers

Now — immediate work

• Knowledge wiki schema — new bundled schema (knowledge-wiki) alongside general and strict_ost. Lightweight: page type labels (source, concept, synthesis, personal), minimal required frontmatter, provenance tracking, loose linking support
• Register a wiki space — add an existing knowledge directory to config using the new schema, let validation surface what needs tidying
• Ingest skill — new Claude Code plugin skill/agent skill for processing raw sources into wiki pages (read source → extract key info → create/update wiki pages → validate)
• Refine schema from real content — evolve the schema by applying it to an existing knowledge base and seeing what works/doesn't

Soon — follow-up work

• index.md awareness — skills or hooks that can detect when pages are missing from an index and prompt the agent to update it (index itself not necessarily a structured-context concern, but the tooling can flag gaps)
• Lint/wiki health skill — validate link integrity, flag orphan pages, detect stale claims, suggest missing cross-references
• Wiki query skill — search wiki pages → synthesize answer → optionally file answer back as new page → validate
• Log.md conventions — optional structured logging format that tools can parse (Karpathy's append-only chronological record)
• Source provenance rules — validate that concept/synthesis pages reference their contributing sources
• Dataview-compatible frontmatter — ensure schema frontmatter works with Obsidian Dataview plugin for dynamic tables

Explore — uncertain/later

• Search integration — at what scale does index.md become insufficient? Hook points for qmd or similar
• Page importance/decay — prevent wiki bloat from filing every query answer
• Contradiction detection — cross-entity coherence rules between pages
• Ingest batch vs interactive — support both supervised single-source ingest and less-supervised batch processing
• Marp/presentation output — generate slide decks from wiki content
• Dashboard for knowledge spaces — overview page reflecting the space's content

Key design principles

• Light enforcement — knowledge wikis should feel organic, not bureaucratic. Schema validates structure, not meaning
• Arbitrary linking — schema covers typed relationships, but [[wikilinks]] between any pages are always valid (no "dangling link" errors for arbitrary cross-references)
• Separation of concerns — schema (structured-context) handles mechanical validation; agent instructions (AGENTS.md, skills) handle behavioural conventions
• Distributable skills — new skills should be reusable by anyone with a knowledge wiki space, not tied to a specific vault setup

Reference

• Karpathy's LLM Wiki gist
• Current bundled schemas: general.json, strict_ost.json