Vault Wiki

Multi-Domain LLM Wiki (TurboVault)

Build and maintain a persistent, compounding federation of knowledge bases as interlinked markdown files. Based on Andrej Karpathy's LLM Wiki pattern, extended for multiple domains.

Unlike traditional RAG (which rediscovers knowledge from scratch per query), each wiki compiles knowledge once and keeps it current. Cross-references are already there. Contradictions have already been flagged. Synthesis reflects everything ingested. Multiple focused wikis live under one root — each with its own schema, index, and log — linked across domain boundaries via path-based wikilinks.

Division of labor: The human curates sources and directs analysis. The agent summarizes, cross-references, files, and maintains consistency across domains.

Storage: All wiki content lives in the wiki/ directory within an Obsidian vault managed by TurboVault. All operations use mcp_turbovault_* tools. See the turbovault-use skill for tool-level mechanics (read/write/edit/search/batch).

When This Skill Activates

Use this skill when the user:

• Asks to create, build, or start a wiki
• Asks to ingest, add, or process a source into their wiki
• Asks a question and an existing wiki exists in the active vault's wiki/ directory
• Asks to lint, audit, or health-check their wiki
• References their wiki, knowledge base, or "notes" in a research context
• Mentions a specific domain wiki by name (e.g., "add this to my ai-research wiki")
• Asks to create a new domain wiki

Wiki Location

The wiki lives at wiki/ within the active TurboVault vault. Set the active vault at the start of each session:

mcp_turbovault_set_active_vault(name="<vault-name>")

All paths in this skill are relative to the vault root with a wiki/ prefix.

Example paths:

• Hub: wiki/index.md
• Schema: wiki/llm-wiki/SCHEMA.md
• Entity page: wiki/llm-wiki/entities/transformer-architecture.md

Architecture

vault-root/                   # active TurboVault vault
├── wiki/                     # wiki root — hub index.md lives here
│   ├── index.md              # Hub: one section per domain wiki with abstract
│   ├── llm-wiki/             # Domain Wiki 1
│   │   ├── SCHEMA.md         # Conventions, structure rules, domain config
│   │   ├── llm-wiki.md       # Sectioned content catalog with one-line summaries
│   │   ├── log.md            # Chronological action log (append-only, rotated yearly)
│   │   ├── raw/              # Layer 1: Immutable source material
│   │   │   ├── articles/     # Web articles, clippings
│   │   │   ├── papers/       # PDFs, arxiv papers
│   │   │   ├── transcripts/  # Meeting notes, interviews
│   │   │   └── assets/       # Images, diagrams referenced by sources
│   │   ├── entities/         # Entity pages (people, orgs, products, models)
│   │   ├── concepts/         # Concept/topic pages
│   │   ├── comparisons/      # Side-by-side analyses
│   │   └── queries/          # Filed query results worth keeping
│   ├── ai-research/          # Domain Wiki 2
│   └── ...                   # Additional domain wikis
├── area/
├── projects/
└── ...

Every level uses a different naming: the hub wiki/index.md catalogs all domain wikis, while each domain wiki has its own <name>-wiki.md (e.g., wiki/llm-wiki/llm-wiki.md). Each domain wiki's full structure: SCHEMA.md, <name>-wiki.md, log.md, raw/, entities/, concepts/, comparisons/, and queries/.

Layer 1 — Raw Sources: Immutable by default. The agent reads but does not modify these during normal ingest/synthesis. If the user explicitly asks to correct or align a raw source itself, treat it as an intentional raw-source revision: make the narrow change and append a log.md entry documenting the exception. Layer 2 — The Wiki: Agent-owned markdown files. Created, updated, and cross-referenced by the agent. Layer 3 — The Schema: Each domain wiki's SCHEMA.md defines structure, conventions, and tag taxonomy for that domain.

Domain wikis are flat subdirectories directly under wiki/ — no wiki/wikis/ nesting.

The Hub

The root wiki/index.md is the hub. It contains one section per domain wiki.

Format: Each section has a ## name heading matching the wiki's directory name, followed by an abstract paragraph describing the wiki's purpose and scope.

# Wiki Hub

## [[wiki/llm-wiki/llm-wiki|llm-wiki]]
LLM-Wiki methodology — persistent, compounding knowledge base for developing the multi-domain LLM Wiki skill itself.

## [[wiki/ai-research/ai-research-wiki|ai-research]]
AI/ML research, models, papers, benchmarks.

Rules:

• No page counts, no last-update dates, no tag lists — those live in each wiki's own <name>-wiki.md and log.md
• No separate routing file — the abstracts are the routing information
• The hub lives in the vault, not in the skill config
• The skill reads it for routing and updates it when creating new domain wikis

Hub detection: The root wiki/index.md is always a hub. If it doesn't exist, the wiki root is uninitialized. → Trigger: Load ./references/initialize-wiki.md when no wiki/index.md exists, or when the user asks to create a new domain wiki.

Routing

How the agent decides which wiki to use:

The agent reads the abstracts in wiki/index.md and matches against them. No tag system, no keyword mapping — just the abstract text.

Resuming an Existing Wiki (CRITICAL — do this every session)

When the user has an existing wiki, always orient yourself before doing anything:

① Ensure the active vault is set: mcp_turbovault_set_active_vault(name="<vault-name>")

② Read root wiki/index.md — it is always a hub. mcp_turbovault_read_note(path="wiki/index.md")

③ Route to target wiki — explicit naming → abstract match → ask (see Routing above).

④ Read that wiki's SCHEMA.md, <wiki>-wiki.md, and recent log.md:

mcp_turbovault_read_note(path="wiki/<target-wiki>/SCHEMA.md")
mcp_turbovault_read_note(path="wiki/<target-wiki>/<target-wiki>-wiki.md")
mcp_turbovault_read_note(path="wiki/<target-wiki>/log.md")

Only after orientation should you ingest, query, or lint. This prevents:

• Creating duplicate pages for entities that already exist
• Missing cross-references to existing content
• Contradicting the schema's conventions
• Repeating work already logged
• Targeting the wrong domain wiki

For large wikis (100+ pages), also run mcp_turbovault_search(query="<topic>") and filter results to wiki/<target-wiki>/ before creating anything new.

Linking

ALL wikilinks must use the full vault path: [[wiki/<wiki-name>/<type>/<page-name>]]

This is a mandatory rule, not a convention — without the wiki/ prefix, wikilinks resolve relative to the current file's directory and TurboVault's broken-link detection will flag every link as broken. Even links within the same domain wiki must use the full path.

Examples:

• [[wiki/ai-research/concepts/llm-infrastructure]] — same-wiki link
• [[wiki/llm-wiki/concepts/three-layer-architecture]] — cross-wiki link
• [[wiki/ai/entities/andrei-karpathy|Andrej Karpathy]] — with display text

Never write [[ai-research/concepts/foo]] or [[concepts/bar]]. These look for pages at the vault root, which doesn't exist. This is the #1 cause of initial lint failures on new wikis.

Additional rules:

• Links resolve natively in Obsidian (backlinks, graph view, manual browsing)
• No adapter pages — link directly using the full path
• No duplication — link, don't copy content between wikis

Initializing a New Wiki

Load and follow references/initialize-wiki.md precisely, in order, every step. All steps are mandatory — especially hub registration. A wiki not registered in the hub is invisible to the routing system.

• Scenario A: First-Time Setup (no hub exists)
• Scenario B: New Domain Wiki (hub already exists)

Core Operations

1. Ingest

→ Trigger: Load ./references/ingest-workflow.md when the user provides a source (URL, paste, file) to integrate into a wiki, or when processing items from vault inbox.

Key rules that apply at every ingest:

• Check for duplicates first — search the target wiki before creating new pages
• Every page needs 2+ cross-references — isolated pages are invisible
• Update <name>-wiki.md and log.md — skipping this degrades the wiki
• Handle truncated content: If web_extract returns a summary instead of full text, label the source status: incomplete and ask the user for the full version. Never create entity/concept pages from truncated sources.

2. Query

When the user asks a question about the wiki's domain:

① Read root wiki/index.md — identify which wiki(s) are relevant using the hub abstracts. ② Read the relevant wiki <name>-wiki.md files to identify relevant pages. ③ For wikis with 100+ pages, also run mcp_turbovault_search(query="<topic>") across the wiki prefix — the index alone may miss relevant content. ④ Read the relevant pages using mcp_turbovault_read_note(path="wiki/..."). ⑤ Synthesize an answer from the compiled knowledge. Cite pages with their vault path: "Based on [[wiki/ai-research/concepts/transformer-architecture]] and [[wiki/llm-wiki/concepts/three-layer-architecture]]..." ⑥ File valuable answers back — if the answer is a substantial comparison, deep dive, or novel synthesis, create a page in the target wiki's queries/ or comparisons/. Don't file trivial lookups — only answers that would be painful to re-derive. ⑦ Update that wiki's log.md with the query and whether it was filed.

3. Lint

→ Trigger: Load ./references/lint-workflow.md when the user asks to lint, audit, or health-check the wiki. Covers 18 checks: broken links, outbound link count, frontmatter validation, index completeness, orphans, tag taxonomy, stale content, page size, log rotation, report, fix cascading, cross-wiki links, hub drift, and source drift (sha256).

Lint cascades — fixing one page often reveals the next. Re-run checks after each fix pass until clean.

Working with the Wiki

Searching

See the turbovault-use skill for available search tools.

Scope warning: TurboVault search searches the entire vault, not just wiki/. Always scope results by wiki/<domain>/ path prefix or use exclude_paths.

Detecting Unprocessed Raw Sources

→ Trigger: Load ./references/detect-unprocessed-sources.md when the user asks to process new articles in raw/articles/. The raw/articles/ directory is the signal — any file there not yet referenced by a wiki page is pending processing.

Using edit_note for Targeted Edits

See the turbovault-use skill for syntax and format requirements.

vault-wiki-specific DON'Ts:

• DON'T use edit_note for log.md — SEARCH replaces the previous entry's header, leaving detail lines orphaned. Always use read-full/write-full.
• DON'T use edit_note for <name>-wiki.md — matching ## Entities deletes the header. Always use read-full/write-full.
• DON'T use edit_note for SCHEMA.md — pipe characters |, brackets [], and backticks trigger parser errors. Always use read-full/write-full.
• DON'T use edit_note for raw/ files — raw sources are immutable by default. Use full read + write for corrections.

Archiving

→ Trigger: Load ./references/archiving.md when the user asks to archive a wiki page or delete an entire domain wiki.

Source Freshness Check

→ Trigger: Load ./references/source-freshness-check.md when the user asks to check if git-based sources (tools, repos) have been updated since ingest. Re-extracts the URL and compares sha256 against stored value. User decides on re-ingest.

Source Cascade Removal

→ Trigger: Load ./references/source-cascade-removal.md when the user asks to remove a raw source and all pages derived from it. Depends on sources: frontmatter for attribution — pages without it need user review.

Self-Improvement Gate

This skill is the stable core. Do not edit it.

All optimizations, pitfalls, and discovered workflows belong in vault-improvements — loaded via the /vault bundle alongside this skill.

When a lesson learned emerges:

1. Do NOT edit this file or its original references — they are the stable core
2. Instead, write the finding into the vault-improvements skill as a new reference entry or section
3. Abstract properly: remove proper names, local paths, session dates, one-off tool names before writing
4. If a lesson is universal and user-approved, it may later be promoted into this skill — but the agent never promotes unilaterally

Do's and Don'ts

• Never modify raw/ files during normal ingest — sources are immutable. Corrections go in wiki pages. Exception: user explicitly asks for a raw-source revision.
• Always orient first — read hub → target SCHEMA + <name>-wiki + recent log before any operation in a new session.
• Always update <name>-wiki.md and log.md — these are the navigational backbone. Skipping them degrades the wiki.
• Don't write summary pages — one topic = one page. A source about "Wohnwagen-Kauf" should produce feuchtigkeitsschaden.md + gasanlage.md + wohnwagen-marken.md, not one summary document.
• Don't create pages for passing mentions — follow the Page Thresholds in the domain wiki's SCHEMA.md.
• Don't create pages without cross-references — every page must link to at least 2 other pages.
• Don't create adapter pages — link directly via [[wiki/other-wiki/concepts/topic]]. No duplication.
• Don't guess the target wiki — if routing is ambiguous, ask. Wrong wiki is worse than asking.
• Frontmatter is required — it enables search, filtering, and staleness detection.
• Tags must come from the taxonomy — freeform tags decay into noise. Add new tags to SCHEMA.md first.
• Keep pages scannable — split pages over 200 lines. Move detailed analysis to dedicated deep-dive pages.
• Ask before mass-updating — if an ingest touches 10+ existing pages, confirm with the user.
• Rotate the log — when log.md exceeds 500 entries, move it to log-YYYY.md and start fresh.
• Handle contradictions explicitly — note both claims with dates, mark in frontmatter, flag for review.
• Keep hub abstracts current — stale abstracts cause routing failures.
• Annotate financial figures as brutto or netto — never record a bare monetary amount without this annotation.
• Verify personal names and relationships — confirm names and relationships before writing. Never infer from context.
• Lint cascades — re-run checks after each fix pass until clean.
• Raw sources can have frontmatter — check the domain SCHEMA. When in doubt, add it.
• Never use truncated content as a raw source — label it status: incomplete and wait for the full text.