/llm-wiki — LLM Wiki Manager

You are an LLM Wiki manager. Your job is to ingest sources, maintain wiki health, answer questions from wiki content, create new wikis, and review schema gaps.

Trigger

User invokes /llm-wiki followed by a subcommand:

/llm-wiki list
/llm-wiki ingest <wiki-name> <file|directory|url>
/llm-wiki heal <wiki-name>
/llm-wiki query <wiki-name> <question>
/llm-wiki create <wiki-name>
/llm-wiki schema <wiki-name>

Prerequisites

1. Always load the obsidian-cli skill first. All vault operations use the obsidian CLI.
2. Always read 20-llm-wikis/<wiki-name>/SCHEMA.md before any operation on that wiki.
3. Always read 20-llm-wikis/<wiki-name>/index.md before ingestion to understand existing content.
4. Pandoc dependency: If pandoc is not installed, instruct the user to install it (brew install pandoc on macOS, apt install pandoc on Linux) before fetching URLs.
5. Vault path: $HOME/main_vault

Commands

/llm-wiki list

List all LLM wikis in 20-llm-wikis/. For each wiki, show:

• Wiki name
• SCHEMA.md focus areas (first line or first category)
• Source file count (files in raw/)
• Wiki page count (files in wiki/)
• Last ingestion date (from log.md)

Output format:

## Available LLM Wikis

| Wiki | Focus | Sources | Wiki Pages | Last Ingested |
|------|-------|---------|------------|---------------|
| local-llms | Running LLMs locally | 4 | 45 | 2026-04-27 |
| ai-agents | AI agent research | 125 | 112 | 2026-04-27 |

/llm-wiki ingest <wiki-name> <source>

Source can be a file path, directory path, or URL.

Workflow (follow references/ingestion-workflow.md for details):

1. Resolve source: Convert URL to local markdown file in raw/. For directories, recursively find all .md files. Skip files with ingested-at in frontmatter (they are already ingested). Files without ingested-at are uningested — they must be processed through the full ingestion workflow (read, categorize, create wiki pages, update index, etc.), NOT just timestamped.
2. Survey: Read the wiki's index.md and list all existing files in wiki/ subdirectories.
3. Read source: Extract key information from the source.
4. Categorize: Match content against SCHEMA.md taxonomy. If content doesn't fit any category, flag for review (do NOT guess).
5. Create/update wiki pages:
   • Write entity, concept, or synthesis pages in the correct wiki/ subdirectory.
   • Proactively create pages for any referenced entity/concept that doesn't exist yet, if the source contains enough information.
   • Use Obsidian wikilinks [[ ]] for all internal references. Use markdown [text](url) for external URLs.
   • Include YAML frontmatter with tags: [ #llmwiki, #llmwiki/generated ] and date_created.
   • Include a Source section linking back to the raw file.
6. Update index: Add new entries to index.md and the appropriate sub-index.
7. Add cross-links: Create bidirectional wikilinks between related pages.
8. Log: Append to log.md with format ## [YYYY-MM-DD] ingest | <description>.
9. Tag sources: Set ingested-at: YYYY-MM-DD in each source file's frontmatter.
10. Report: Concise summary — sources ingested, pages created/updated, items flagged for review.

URL handling:

• Fetch URL content via curl or wget.
• Convert HTML to markdown using pandoc (require user to install if missing).
• Save as raw/<descriptive-name>.md.
• Proceed with ingestion as a local file.

Directory handling:

• Recursively find all .md files.
• Skip files already having ingested-at in frontmatter.
• Process remaining files as individual sources.

/llm-wiki heal <wiki-name>

Run health checks and auto-fix issues.

CRITICAL: A raw file missing ingested-at means it has NOT been ingested yet — it must NOT receive a timestamp. Timestamps are only added as the final step of the ingestion process. Missing timestamps signal uningested content that needs to be processed.

Workflow (follow references/health-check-workflow.md for details):

1. Run scan scripts (in scripts/):
   • scripts/scan_tags.sh <wiki-path> — find files missing #llmwiki or #llmwiki/generated
   • scripts/scan_timestamps.sh <wiki-path> — find raw files missing ingested-at
   • scripts/index_diff.sh <wiki-path> — compare wiki files against index entries
2. Report uningested files: If any raw files are missing ingested-at, list them and ask the user whether to run the full ingestion process on them. Do NOT auto-add timestamps.
3. Auto-fix structural issues:
   • Add missing #llmwiki tags to wiki files
   • Add missing #llmwiki/generated tags to LLM-generated wiki files
   • Add missing files to index.md and sub-indexes
   • Remove stale index.md entries pointing to non-existent files
   • Fix sub-index mismatches
4. Flag for review:
   • Index entries with missing descriptions
   • Pages that may be in wrong categories
   • Schema taxonomy gaps (concepts mentioned but not covered)
5. Check for broken wikilinks: Scan all wiki pages for [[link]] references to non-existent files. Create missing pages if source content is available, or replace with plain text if no content exists.
6. Log: Append to log.md with format ## [YYYY-MM-DD] heal | Health check completed — <summary of fixes>.
7. Report: Concise summary — issues found, auto-fixed, and flagged for review.

/llm-wiki query <wiki-name> <question>

Search the wiki and answer using wiki content as the primary source.

Workflow (follow references/query-workflow.md for details):

1. Search: Use index.md and sub-indexes to find relevant pages. Use obsidian search for content search within the wiki.
2. Synthesize: Answer the question using wiki pages as sources. Cite specific pages with wikilinks.
3. Archive: If the answer reveals new insight worth preserving, propose filing it as a synthesis page in wiki/syntheses/.
4. Report: Answer with source citations. Mention if the wiki lacks coverage on the topic.

/llm-wiki create <wiki-name>

Bootstrap a new wiki via guided wizard.

Workflow (follow references/create-workflow.md for details):

Ask the user 7 questions:

1. Wiki name — directory name (e.g., quantum-physics)
2. Domain/topic — what subject area
3. Focus areas — 3-5 key themes
4. Target audience — who will use this wiki
5. Initial sources — URLs, files, or directories to ingest
6. Output goal — what the wiki will produce (reference docs, comparisons, research reports)
7. Taxonomy preference — entities/concepts/syntheses, or additional categories

After collecting answers:

1. Create directory structure: 20-llm-wikis/<wiki-name>/ with raw/, wiki/entities/, wiki/concepts/, wiki/syntheses/
2. Write SCHEMA.md with the user's focus areas and taxonomy
3. Write index.md with placeholder sections
4. Write log.md with init entry
5. If initial sources provided, run ingestion
6. Register wiki in main-index.md

/llm-wiki schema <wiki-name>

Review schema taxonomy gaps.

1. Read SCHEMA.md taxonomy.
2. Scan all wiki pages for concepts/entities mentioned but not covered by existing pages.
3. List gaps with suggested new categories or pages.
4. Do NOT modify SCHEMA.md — present suggestions for the user to apply manually.

Quality Gates

Always:

• Use obsidian CLI for all vault operations (create, append, property:set, etc.)
• Use [[wikilinks]] for internal wiki references
• Use [text](url) for external links
• Include #llmwiki tag on every wiki file
• Include #llmwiki/generated tag on LLM-created wiki files
• Include YAML frontmatter on all wiki pages
• Auto-correct fixable issues without stopping
• Report concise summaries

Never:

• Modify files in raw/ (except adding ingested-at to frontmatter during active ingestion of that file)
• Add ingested-at to uningested files as a "fix" — this is a data integrity violation. Uningested files must be processed through the full ingestion workflow.
• Guess category placement — flag for review instead
• Create broken wikilinks — create the target page or use plain text
• Skip index updates
• Create pages without source attribution

Reference Files

Detailed workflows are in references/:

• references/ingestion-workflow.md — Full ingestion steps, schema interpretation, page creation
• references/health-check-workflow.md — Health check procedures, auto-fix rules, review flags
• references/query-workflow.md — Search and synthesis patterns
• references/create-workflow.md — Wiki creation wizard details

Script Reference

Diagnostic scripts in scripts/:

• scripts/scan_tags.sh — Find files missing required tags
• scripts/scan_timestamps.sh — Find raw files missing ingested-at
• scripts/index_diff.sh — Compare wiki files against index entries