Query

Skill: /query

Trigger

User types: /query

STRICT RULE

NEVER read from raw/ during /query. raw/ is the source archive — only /compile reads it. wiki/ is the LLM-readable knowledge base — /query ALWAYS reads from wiki/ only. If a topic has not been compiled into wiki/ yet, say so and suggest running /ingest.

Step 1 — ROUTE (read vocabs/index.md)

1. Read the FULL vocabs/index.md file (~200 tokens)
2. From the question, identify which vocab categories are relevant Example: "When should I use Redis vs Kafka?" → matches categories: technologies, data-storage, system-design
3. Note the canonical term names from those categories (these are the tags you will match against in Step 2)

Step 2 — FILTER (read wiki/index.md)

1. Read wiki/index.md
2. Scan EVERY section for entries whose tags overlap with Step 1 terms:

   • Concepts

   • Entities

   • Patterns

   • Best Practices

   • Comparisons

   • Sources

   • Saved Queries ← DO NOT SKIP THIS SECTION

3. From all matching entries across all sections, select top 3-7 candidates ranked by how many tags overlap with the question's terms

Priority Rule for Saved Queries:

If a "## Saved Queries" entry has tags that match the question: → ALWAYS include that page as a candidate, ranked FIRST → A saved query was compiled specifically to answer a question like this → It will likely satisfy the query with far fewer tokens than reading raw pages

Step 3 — READ (open only matched pages)

For each candidate page (starting from top-ranked):

1. Open the page

2. Read ONLY the > blockquote summary line (first line after frontmatter):

   • If summary directly answers or closely matches the question: → Read the full page → If this is a Saved Query page AND the answer is sufficient: STOP here. Do not open remaining candidates. Go to Step 4.
   • If summary does not match: → Skip this page, move to next candidate

3. Continue until you have 2-5 pages with confirmed relevant content

Short-circuit rule:

If a Saved Query page fully answers the question → skip all remaining candidates. No need to re-read the underlying concept/entity pages. This is the token-saving payoff of saved queries.

Step 4 — ANSWER

• Synthesize a clear answer from the pages read
• Cite pages inline: [[page-name]]
• If answer came primarily from a Saved Query page, note that: "Answer sourced from saved query [[page-slug]]"
• State confidence: high | medium | low
• State which pages were consulted and how many tokens were saved vs. reading from scratch (estimate: "read 1 saved query vs ~5 pages")

Step 5 — SAVE DECISION

Ask the user: "Worth saving to wiki? (y/n)"

If no → stop.

If yes → do ALL of the following:

5a — Write the Query Page

File: wiki/queries/.md Template from wiki/CLAUDE.md (Query Page template): Frontmatter: t: query cat: <most relevant vocab category from Step 1> tags: [] src: 0 rel: [<all pages consulted in Step 3>] upd: Opening line: > Answer to: "" Sections: ## Answer ← full synthesized answer with [[citations]] ## Pages Consulted ← [[page1]], [[page2]] as wikilinks (for graph edges) ## Confidence ← high | medium | low ## Follow-up Questions ← related questions worth exploring

5b — Update wiki/index.md

Under "## Saved Queries" section, add:

• [[]] | query | tag1,tag2,tag3 | 0 (tags must match canonical vocab terms so future queries can find this) Update frontmatter:
• Increment pages: count by 1
• Update upd: to today's date

5c — Update wiki/log.md

Append:

[YYYY-MM-DD] query |

5d — Backlink from Consulted Pages (prevents orphan in graph)

For each wiki page read in Step 3:

• Open the page
• Under its "## Sources" section add:
  • [[]]: referenced in query ""
• Increment its src: frontmatter count by 1

This ensures the saved query page is connected in the Obsidian graph and will NOT be flagged as an orphan by /lint.