aksheyw LLMwiki — Persistent Project Knowledge Wiki
A file-based wiki system that captures durable project knowledge across Claude Code sessions. Instead of re-deriving context from raw session files each time, the wiki maintains structured, searchable, cross-referenced markdown pages that compound over time.
Credit: the core idea (an LLM incrementally building and maintaining a persistent, interlinked wiki that grows richer as knowledge is ingested) is Andrej Karpathy's LLMwiki pattern, from his gist. This skill is an implementation of that pattern for Claude Code.
When to Use
- After fixing a non-obvious bug — capture the root cause so future sessions don't re-investigate
- After making an architectural decision — record the reasoning, not just the choice
- After a session with significant learnings — extract the 3-5 durable insights before they're buried in session data
- When starting a new session — query the wiki instead of re-reading 30MB of
.jsonlfiles - Periodically — lint the wiki to catch contradictions with current code/docs
Commands
| Command | Purpose |
|---|---|
/wiki-init |
Set up .claude/wiki/ in current project |
/wiki-ingest [source] |
Capture learnings into wiki pages |
/wiki-query <question> |
Search wiki and synthesize an answer |
/wiki-lint |
Audit wiki for staleness, contradictions, orphans |
Wiki Structure (per project)
<project-root>/.claude/wiki/
_index.md # Topic catalog (the brain — kept under 200 lines)
_log.md # Chronological audit trail of all changes
_schema.md # Project-specific topic categories
<topic>.md # Individual knowledge pages
How It Complements Existing Systems
The systems named below are examples from the author's setup, not dependencies of this skill — the point is where a wiki fits among whatever learning/memory systems you run:
| System | Scope | LLMwiki Relationship |
|---|---|---|
| Instinct/lesson capture (e.g. continuous-learning loops) | Atomic instincts (trigger→action) | Wiki stores the reasoning behind instincts |
| Session narratives (e.g. save-session / resume-session commands) | Full session narratives | Wiki ingests from sessions, extracts durable facts |
| MEMORY.md | Cross-project strategic decisions | Wiki handles project-specific knowledge |
| docs/ (INDEX.md) | Designed architecture docs | Wiki handles emergent knowledge discovered during work |
Key Principles
- Never duplicate — always merge new information into existing pages
- Cross-reference with Obsidian-style
[[links]]— every page uses[[page-name]]syntax so the wiki renders as a graph in Obsidian - YAML frontmatter on every page —
created,updated,source,tags,aliases(enables Obsidian queries and filtering) - Date everything — each fact includes when it was learned
- Source everything — cite the session, file, or conversation where knowledge came from
- Keep _index.md lean — one line per topic, max 200 lines. If it grows beyond that, reorganize into sub-topics.
- Manual ingest only — the user/agent decides what's worth capturing. Auto-capture creates noise.
- Don't preload at session start — use
/wiki-queryon demand to avoid context bloat. - Obsidian-compatible — the wiki should be openable as an Obsidian vault (File → Open folder as vault →
.claude/wiki/) with graph view working out of the box.
Operations Reference
See references/operations.md for detailed operation specs.
See references/schema-template.md for the default wiki schema.