wiki（wiki 搜索）

LLM Wiki

Persistent, compounding knowledge base inside an Obsidian vault.

Operations

/llm-wiki:wiki init my-topic
/llm-wiki:wiki ingest ~/ObsidianVault/03-Resources/my-topic/raw/article.md
/llm-wiki:wiki ingest https://example.com/article
/llm-wiki:wiki query "What is X?"
/llm-wiki:wiki lint

Active Wiki Detection

Walk up from cwd looking for a directory containing both CLAUDE.md and a wiki/ subfolder.

1. Start at cwd. Check if CLAUDE.md and wiki/ exist in the current directory.
2. If found → that directory is the active wiki root. Read CLAUDE.md for schema.
3. If not found → move to parent directory and repeat until filesystem root.
4. If no wiki found anywhere in the path, prompt the user: "Which wiki should I use?" List available wikis by running: ls -d ~/ObsidianVault/03-Resources/*/wiki 2>/dev/null and presenting the parent directory names.

qmd Availability

Reference paths used throughout this skill:

QMD="env -u BUN_INSTALL ${CLAUDE_PLUGIN_DATA}/node_modules/.bin/qmd"
MARP="${CLAUDE_PLUGIN_DATA}/node_modules/.bin/marp"

Check: Test if "${CLAUDE_PLUGIN_DATA}/node_modules/.bin/qmd" exists and is executable via Bash: test -x "${CLAUDE_PLUGIN_DATA}/node_modules/.bin/qmd".

Important: Always invoke qmd via env -u BUN_INSTALL to force Node.js runtime. If BUN_INSTALL is set in the environment, qmd runs under Bun, which uses a SQLite build without extension loading support and cannot load sqlite-vec.

• If present: use it for query and embed operations. ALWAYS use the full path — never bare qmd.
• If absent: fall back to reading wiki/index.md manually and grepping wiki files.

init <name>

Create a new wiki scaffold under the Obsidian vault.

Steps

1. Check if wiki already exists: If ~/ObsidianVault/03-Resources/<name>/ exists, abort with: "Wiki '' already exists at ~/ObsidianVault/03-Resources//. Use wiki remove <name> first, or choose a different name."
2. Create directory structure: mkdir -p ~/ObsidianVault/03-Resources/<name>/raw/articles mkdir -p ~/ObsidianVault/03-Resources/<name>/raw/attachments mkdir -p ~/ObsidianVault/03-Resources/<name>/wiki/queries mkdir -p ~/ObsidianVault/03-Resources/<name>/outputs/reports
3. Write ~/ObsidianVault/03-Resources/<name>/CLAUDE.md using the CLAUDE.md template below (fill in <name>).
4. Write ~/ObsidianVault/03-Resources/<name>/wiki/index.md using the index.md template below.
5. Write ~/ObsidianVault/03-Resources/<name>/log.md using the log.md template below.
6. Write ~/ObsidianVault/03-Resources/<name>/.gitignore using the .gitignore template below.
7. Write ~/ObsidianVault/03-Resources/<name>/qmd.yml using the qmd.yml template below.
8. Commit to vault git: git -C ~/ObsidianVault add "03-Resources/<name>/" && git -C ~/ObsidianVault commit -m "init: <name> wiki"
9. If qmd available: "${QMD}" collection add ~/ObsidianVault/03-Resources/<name>/wiki --name <name> && "${QMD}" embed --collection <name>
10. Print Web Clipper setup instruction: Obsidian Web Clipper setup: 1. Install: https://obsidian.md/clipper 2. In clipper settings, set Destination folder to: 03-Resources/<name>/raw/articles 3. Set filename template to: {{date:YYYY-MM-DD}}-{{title}} 4. After clipping, run: /llm-wiki:wiki ingest ~/ObsidianVault/03-Resources/<name>/raw/articles/<clipped-file>.md

ingest <path|url>

Acquire a source and save it to the raw library. Does NOT create wiki pages — use compile for that.

Steps

1. Detect active wiki (see Active Wiki Detection). Read CLAUDE.md for schema.
2. Acquire source:

- If input is a URL: use the WebFetch tool to retrieve content. - If input is a file path: read the file directly. - If the source file is in raw/ directly (not in a subdirectory), read it from there. Sources saved before the raw/articles/ convention are still valid.

1. Classify the source as one of: article | paper | transcript | conversation | image-set.
2. Save to raw library: Write to raw/articles/YYYY-MM-DD-<slug>.md with frontmatter: --- date: YYYY-MM-DD source-type: <classification> source-url: <original URL or file path> title: <extracted or inferred title> compiled: false --- If input was a file path already in raw/, skip this step (source is already saved).
3. Append to log.md: ## [YYYY-MM-DD] ingest | <title> Saved <source-type> from <source> to raw/articles/.
4. Commit: git -C ~/ObsidianVault add "03-Resources/<wiki-name>/" && git -C ~/ObsidianVault commit -m "ingest: <title>"
5. Print: "Source saved to raw/articles/. Run wiki compile to integrate into the wiki."

compile [<path>]

Read raw sources and create/update wiki pages with entity extraction and cross-references.

• If <path> is given: compile that specific raw source.
• If no argument: scan raw/articles/ for sources without a corresponding source-summary page in wiki/, and compile those.

Steps

1. Detect active wiki. Read CLAUDE.md for schema and templates.
2. Identify sources to compile:

- If path argument given: use that file. - Otherwise: list files in raw/articles/. For each, check if a corresponding source-summary page exists in wiki/ (match by slug or title). Compile any source without a matching summary. - If nothing to compile: "All sources are already compiled. Nothing to do." Stop.

1. For each source to compile: a. Read the raw source content. b. Write or update source-summary page in wiki/ using the source-summary template from CLAUDE.md. Filename: <slug>.md. c. Entity extraction: For each mentioned entity (person, concept, event): d. Backlink audit (CRITICAL — do not skip): grep -rln "<new page title>" wiki/ For each file that mentions the new page title but does NOT contain [[new-page-name]]: add a [[wikilink]] at the first mention.

- Check if a page already exists in wiki/. - If yes → update with new information, preserving existing content. - If no → create using the appropriate template (concept.md or person.md). - Add [[wikilinks]] to related pages in both directions.

1. Update wiki/index.md with new/updated entries under the appropriate domain heading.
2. Append to log.md: ## [YYYY-MM-DD] compile | <N> sources → <M> pages Compiled <source-titles>. Created/updated M pages.
3. Commit: git -C ~/ObsidianVault add "03-Resources/<wiki-name>/" && git -C ~/ObsidianVault commit -m "compile: <summary>"
4. If qmd available: "${QMD}" embed --collection <name>

query <question>

Answer a question using wiki knowledge, with citations.

Steps

1. Detect active wiki. Read CLAUDE.md.
2. Find relevant pages:

- If qmd available: "${QMD}" query "<question>" --collection <name> Parse output for candidate page paths. - Otherwise: read wiki/index.md and identify relevant pages by title/description matching.

1. Read all relevant pages. Follow one level of [[wikilinks]] if targets look relevant to the question.
2. Synthesize answer with [[wikilinks]] as citations. Format rules:

- Default: prose with inline wikilink citations. - If question contains "table": markdown table with wikilink citations in cells. - If question contains "slides": Marp markdown with marp: true frontmatter. Render with: "${MARP}" <file> -o output.html

1. File the answer to wiki/queries/<slug>.md using the query-output frontmatter schema. Always file — no prompt.
2. Ask: "Promote this answer to wiki/<slug>.md as a concept page? (y/n)"

- If yes: move the file from wiki/queries/ to wiki/, update frontmatter status from filed to promoted, and append to log.md: ## [YYYY-MM-DD] promote | <slug> Promoted query answer to concept page.

1. Append to log.md: ## [YYYY-MM-DD] query | <question-slug> Answered question. Referenced N pages. Filed to queries/<slug>.md.
2. Commit: git -C ~/ObsidianVault add "03-Resources/<wiki-name>/" && git -C ~/ObsidianVault commit -m "query: <slug>"

lint

Audit wiki integrity and fix issues.

Steps

1. Read all files in wiki/.
2. Build a link graph: for each [[wikilink]] on each page, record the edge (source → target).
3. Run deterministic lint script if available: python3 "${CLAUDE_PLUGIN_ROOT}/scripts/lint-wiki.py" <wiki-root>/wiki/
4. Report and fix: Check Action Orphan pages (no inbound links) List them. Suggest adding links from related pages. Dead links ([[wikilinks]] to nonexistent files) Create stub pages with appropriate template. Unlinked concept mentions Scan pages for proper nouns and technical terms appearing in prose without [[wikilinks]]. If a corresponding page exists, add the wikilink at first mention. If no page exists, flag as a candidate for a new page. Contradictions Scan for [!WARNING] markers. List them. Missing "Counter-Arguments and Gaps" sections Add empty ## Counter-Arguments and Gaps section. Stale pages Flag pages with status: stale in frontmatter. Index drift Compare index.md entries vs actual files. Add missing, remove dead.
5. Suggest growth opportunities: Based on the wiki's current content and the gaps found in step 4, generate:

- 3-5 questions the wiki cannot yet answer well (candidates for wiki query) - 2-3 topic areas or sources that would most strengthen the wiki

1. Write lint report to outputs/reports/YYYY-MM-DD-lint.md: # Lint Report — YYYY-MM-DD **Wiki:** <name> | **Issues found:** N | **Fixed:** M ## Issues <full issue table from step 4> ## Next Steps <growth suggestions from step 5>
2. Append to log.md: ## [YYYY-MM-DD] lint | N issues found, M fixed <summary of issues>
3. Commit: git -C ~/ObsidianVault commit -am "lint: YYYY-MM-DD"

remove <name>

Delete a wiki and all its contents.

Steps

1. Resolve wiki path: ~/ObsidianVault/03-Resources/<name>/
2. Verify it exists. If not, abort: "Wiki '' does not exist."
3. Confirm with user: List the directory contents and ask "This will permanently delete the '' wiki and all its contents. Proceed? (y/n)"
4. Remove qmd collection (if qmd available): "${QMD}" collection remove <name>
5. Remove from git and filesystem: git -C ~/ObsidianVault rm -rf "03-Resources/<name>/" && git -C ~/ObsidianVault commit -m "remove: <name> wiki"
6. Confirm: "Wiki '' has been removed."

Error Handling

Handle these failure modes gracefully:

Templates

Used by the init operation. Apply verbatim, replacing <name> with the wiki name.

CLAUDE.md

# <name> Wiki Schema

## Directory Layout
- raw/              -- immutable source drops. Never edit files here.
- raw/articles/     -- text source documents (articles, papers, transcripts).
- raw/attachments/  -- images and binary attachments.
- wiki/             -- LLM-owned pages. You have full write access here.
- wiki/index.md     -- catalog. Read this FIRST before opening any other page.
- wiki/queries/     -- filed query answers. Promote to wiki/ when durable.
- outputs/reports/  -- dated lint reports and other artifacts.
- log.md            -- append-only operation log. Never edit existing entries.

## Entity Types and Templates

### concept.md
---
date: YYYY-MM-DD
tags: [domain]
type: concept
status: active
---
# Concept Name
<one-paragraph summary>

## Details
...

## See Also
- [[related-concept]]

## Counter-Arguments and Gaps
...

### person.md
---
date: YYYY-MM-DD
tags: [domain, person]
type: person
status: active
---
# Person Name
Role / affiliation.

## Key Contributions
...

## See Also
- [[related-concept]]

### source-summary.md
---
date: YYYY-MM-DD
tags: [domain]
type: source-summary
source-url: https://...
---
# Source Title
One-paragraph summary.

## Key Points
...

## Entities Mentioned
- [[person-or-concept]]

## Slides
To export as a Marp slide deck, add `marp: true` to frontmatter and run:
  "${MARP}" wiki/<filename>.md -o output.html

### query-output.md
---
date: YYYY-MM-DD
tags: [domain]
type: query-output
question: "<original question>"
status: filed
---
# <Question as title>
<synthesized answer>

## Sources
- [[page-1]]
- [[page-2]]

## Naming Conventions
- All filenames: lowercase-kebab-case.md
- Wikilinks: [[filename-without-extension]]
- Never use standard markdown links for internal links

## Log Format
Append to log.md after every operation. Format:
  ## [YYYY-MM-DD] <operation> | <title>
  <one-line description>

Operations: ingest | compile | query | lint | promote | remove

## Index Format
wiki/index.md is a human- and LLM-readable catalog. Format:
  ## Domain Name
  - [[page-name]] -- one-line description (YYYY-MM-DD)

Keep entries under 80 chars. Update after every ingest.

## Cross-Reference Rules
- Every page must link to at least one other page when content warrants it
- When creating or updating a concept page, scan index.md for related entities and add [[wikilinks]]
- Flag contradictions inline: > [!WARNING] Contradiction with [[other-page]]

## Ingest Rules
1. Acquire the source (URL or file)
2. Classify the source type
3. Save to raw/articles/ with frontmatter (compiled: false)
4. Append to log.md
5. Commit
Ingest does NOT create wiki pages. Use compile for that.

## Compile Rules
1. Identify uncompiled raw sources (no matching source-summary in wiki/)
2. For each source: write source-summary, extract entities, create/update pages
3. Backlink audit: grep existing pages for mentions of new titles
4. Update wiki/index.md
5. Append to log.md
6. Commit
One source typically touches 5-15 pages. This is normal.

## Query Rules
1. Read wiki/index.md first
2. Open relevant pages
3. Synthesize answer with [[wikilinks]] as citations
4. Always file answer to wiki/queries/ (mandatory, no prompt)
5. Offer promotion to wiki/ as a concept page (y/n)
6. Append to log.md (both query and optional promote events)
7. Commit changes

## Lint Rules
Scan all pages in wiki/ and report:
- Contradictions between pages
- Orphan pages (no inbound [[links]])
- Pages with status: stale older than 90 days
- Missing Counter-Arguments and Gaps section
- Index entries pointing to missing files
After fixing, append to log.md and commit.

.gitignore

.DS_Store
*.sqlite
*.sqlite-wal
*.sqlite-shm

qmd.yml

collections:
  <name>:
    path: ./wiki
    pattern: "**/*.md"

wiki/index.md

# <name> Wiki Index

Last updated: YYYY-MM-DD

<!-- Add entries after each ingest. Format:
## Domain
- [[page-name]] -- description (YYYY-MM-DD)
-->

log.md

# <name> Wiki Log

<!-- Append only. Never edit existing entries. Format:
## [YYYY-MM-DD] ingest | Title
One-line description.
-->