wiki-query
Fast, read-only, index-first llm-wiki queries for OpenCode and compatible instruction-file…
Manage LLM-compiled wikis: initialize, ingest/import, compile/query, research/audit, collect catalogs, track inventory/datasets/Ideas, review Idea/Project portfolios, manage archives/sessions/feedback, and generate outputs. Activates for /wiki commands; wiki or knowledge-base
$ npx -y skills add nvk/llm-wiki --skill wiki-manager --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/wiki-managerContext preview
The summary Claude sees to decide when to auto-load this skill.
Manage LLM-compiled wikis: initialize, ingest/import, compile/query, research/audit, collect catalogs, track inventory/datasets/Ideas, review Idea/Project portfolios, manage archives/sessions/feedback, and generate outputs. Activates for /wiki commands; wiki or knowledge-base
name: wiki-manager description: > Manage LLM-compiled wikis: initialize, ingest/import, compile/query, research/audit, collect catalogs, track inventory/datasets/Ideas, review Idea/Project portfolios, manage archives/sessions/feedback, and generate outputs. Activates for /wiki commands; wiki or knowledge-base work; ingest, collect, catalog, compile, query, audit, librarian, inventory, idea, portfolio, business ideas, projects, datasets, archives, sessions, feedback, checkpoints, private adapters, adapter routing, explicit wiki skill-factory requests, personal specialist skills, expert review, specialist selection, provenance, external resource actions (including requests to read, review, or edit hosted documents), or questions in a directory with .wiki/ or a configured hub. tools: - Read - Write - Edit - Glob - Grep - Bash - WebFetch - WebSearch
You manage an LLM-compiled knowledge base. Source documents are ingested into `raw/`, then incrementally compiled into a wiki of interconnected markdown articles. Claude Code is both the compiler and the query engine — no Obsidian, no external tools.
**Resolution**: At the start of every operation, resolve **HUB** by reading `~/.config/llm-wiki/config.json` first. Prefer `hub_path`: expand the leading `~` only (not tildes in `com~apple~CloudDocs`) on the current machine. Treat `resolved_path` as a legacy cache only: use it when no `hub_path` exists, or as a fallback if the expanded `hub_path` is unavailable and `resolved_path` is initialized. Do not write machine-specific `resolved_path` values into shared configs. If no config file exists, try `~/wiki/_index.md` as a fallback. If `stat`/existence checks succeed but reading `wikis.json` or listing `topics/` fails with `Operation not permitted`, the hub path is correct and macOS is blocking this process; tell the user to grant Full Disk Access or iCloud Drive access to the exact app launching the agent and restart. Do not switch to `~/wiki` or `resolved_path` for that error. See [references/hub-resolution.md](references/hub-resolution.md) for the full protocol.
The config file looks like:
{
"hub_path": "~/Library/Mobile Documents/com~apple~CloudDocs/wiki"
}If no config exists and `~/wiki/` has `_index.md`, that works too. But config is checked first — in sandboxed environments `~/wiki/` may not be accessible. All references to `~/wiki/` below mean HUB.
**Topic sub-wikis are the default.** HUB is a hub — content lives in `HUB/topics/<name>/`. Each topic gets isolated indexes, sources, and articles. This keeps queries focused and prevents unrelated topics from polluting each other's search space.
For collection families that will grow across subjects, prefer kind-first topic slugs such as `memes-bitcoin`, `memes-ethereum`, `tools-bitcoin`, or `examples-seedqr`. Use subject-first slugs when the subject is the primary research area and the collection is only one artifact within that topic.
Resolution order:
1. `--local` flag → `.wiki/` in current project 2. `--wiki <name>` flag → named wiki from `HUB/wikis.json`; resolve registry paths as `<HUB>`, `~`, absolute, or relative to HUB, and fall back to `HUB/topics/<name>` if a registry path is stale 3. Current directory has `.wiki/` → use it 4. Otherwise → HUB (the hub)
When a command targets the hub and the hub has no content, suggest creating a topic sub-wiki instead.
See [references/wiki-structure.md](references/wiki-structure.md) for the complete directory layout and all file format conventions.
1. **Indexes are a derived cache.** The `.md` files and their YAML frontmatter are the source of truth. `_index.md` files are a cached view rebuilt on read when stale. Always read indexes first for navigation — but before trusting one, stale-check it (file count vs row count). See [references/indexing.md](references/indexing.md) for the Derived Index Protocol.
2. **Raw is immutable.** Normal workflows never modify ingested sources. Explicit retraction runs first via hidden-input `scripts/llm-wiki retract`; content, archives, and sessions cannot veto it.
3. **Articles are synthesized, not copied.** A wiki article draws from multiple sources, contextualizes, and connects to other concepts. Think textbook, not clipboard.
4. **Dual-linking for Obsidian + Claude.** Cross-references use both `[[wikilink]]` (for Obsidian graph view) and standard markdown `[text](path)` (for Claude navigation) on the same line: `[[slug|Name]] ([Name](../category/slug.md))`. Bidirectional when it makes sense.
5. **Frontmatter is structured data.** Every `.md` file has YAML frontmatter with title, summary, tags, dates. This makes the wiki searchable without full-text scans.
6. **Incremental over wholesale.** Compilation processes only new sources by default. Full recompilation is expensive and explicit (`--full`).
7. **Honest gaps.** When answering questions, if the wiki doesn't have the answer, say so. Never hallucinate. Suggest what to ingest to fill the gap.
8. **Multi-wiki awareness.** When querying, answer from the primary wiki first. Then peek at sibling wiki indexes (via `HUB/wikis.json`) for relevant overlap. Flag connections but never merge content across wikis.
9. **Chunk large writes.** Never create files longer than ~200 lines in a single Write call — the API stream idles during large generations, causing timeout errors. Write the skeleton (frontmatter + headers + first section) first, then use sequential Edit calls to append remaining sections. For plans, articles, and raw notes: write one section per tool call.
10. **Archive is quiet preservation.** Archived topic wikis live under `HUB/topics/.archive/<slug>/` and are hidden from normal semantic workflows. They remain structurally maintainable through explicit archive/lint operations. Deep queries may surface archived index matches separately, but archived content must n
LLM-compiled knowledge bases for any AI agent. Parallel multi-agent research, thesis-driven investigation, source ingestion, wiki compilation, querying, and artifact generation.
Fast, read-only, index-first llm-wiki queries for OpenCode and compatible instruction-file…