Query
The /query skill lets you ask questions against a vault’s wiki and get synthesized answers with citations. Add --save to file answers back into the wiki — your explorations compound.
/query "What deployment patterns have we seen?" --vault my-research/query "Compare framework A vs B" --vault my-research --save/query "Who are the key people in this space?" --vault my-research| Flag | Description |
|---|---|
--vault <name> | Target vault (uses sole vault if only one exists) |
--save | File the answer back into the wiki as a new page |
How It Works
Section titled “How It Works”- Read the index —
wiki/index.mdis the primary discovery mechanism - Identify relevant pages — source-notes, entities, concepts, comparisons that relate to the question
- Read pages — typically 3-10 pages depending on complexity
- Synthesize — provide a cited answer using
[[wikilinks]] - Save (if
--save) — create a comparison or summary page with full frontmatter
What a Good Answer Looks Like
Section titled “What a Good Answer Looks Like”- Directly answers the question with specifics from the wiki
- Cites sources: “According to [[source-name]], …”
- Notes gaps if the wiki doesn’t fully cover the topic
- Suggests follow-ups — related questions, sources worth ingesting
The --save Flag
Section titled “The --save Flag”When --save is set, the answer becomes a wiki page:
- Comparison questions create pages in
wiki/comparisons/ - General synthesis creates summary pages in
wiki/ - The page gets full frontmatter (
page-type,sources,domain,related) - Index and log are updated, vault is auto-committed
This is the compounding mechanism Karpathy describes: “I end up filing the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always add up.”
- If the wiki is too small to answer meaningfully, query will say so and suggest sources to ingest
- For complex questions, break them into focused sub-questions
- Use
--saveliberally — saved answers become first-class wiki pages that future queries can reference - At large scale (~200+ pages), pair with
/searchfor better page discovery