| name | llm-wiki-query |
| description | Answer a research question by consulting the personal wiki, synthesizing across pages, and producing a cited, cross-linked response. Use when the user asks any domain question, wants to understand a concept, compare models, or trace an idea through the literature — especially when they have an existing wiki. Also handles saving the answer as a new wiki page if the user wants to keep it. Trigger on: "what does the wiki say about", "explain", "compare", "how does X relate to Y", "summarize what we know about", or any open research question in the domain.
|
LLM Wiki — Query Skill
Answer research questions by reading the wiki first, then synthesizing a response with dense cross-references.
Step 1 — Read the Index
Always start by reading wiki/index.md. This gives you the map of everything that has been ingested. Do not answer from memory alone — always ground your response in wiki pages.
Step 2 — Identify Relevant Pages
From the index, identify which pages are relevant to the question. Relevant pages typically include:
- Concept pages for core ideas mentioned in the question
- Paper pages that introduce or evaluate those ideas
- Model pages if the question is architecture-specific
- Author pages only if the question is about a researcher's body of work
List the pages you're about to read before reading them, so the user can see your reasoning.
Step 3 — Read Relevant Pages
Read each identified page in full. As you read, note:
- Direct answers to the question
- Contradictions between pages (e.g., paper A claims X, paper B claims ¬X)
- Gaps — aspects of the question the wiki doesn't address
- Related threads the user might not have asked about but would want to know
Step 4 — Synthesize and Respond
Write the answer using [[wiki-link]] syntax to reference every page you draw from. Structure the response based on the question type:
Concept explanation
- Definition → Intuition → Math → Why it matters → Open questions
- Cite
[[paper_id]] for every non-trivial claim
Model comparison
- Side-by-side on: equivariance type, message-passing order, benchmark results, computational cost
- Note which benchmarks appear in which
[[paper_id]]
Literature trace ("how has X evolved?")
- Chronological structure
- Use
[[paper_id]] citations throughout
- End with current frontier and open problems
Open question / thesis-level question
- Synthesize what the wiki currently supports
- Explicitly flag what is not yet answered
- Suggest which ingested papers are most relevant next steps
Step 5 — Flag Gaps
If the question touches on something the wiki doesn't cover, be explicit:
"The wiki doesn't currently have a page on [topic]. Based on what I've read, [best current answer]. You might want to ingest [related paper/concept] to fill this gap."
Do not fabricate citations. If you're reasoning beyond the wiki, say so clearly.
Step 6 — Offer to Save
After answering, ask:
"Should I save this as a wiki page? If so, I'll add it under wiki/concepts/ or wiki/notes/ and update the index and log."
If the user says yes:
- Choose an appropriate
snake_case page ID
- Create the page with YAML frontmatter (consult
wiki/schema.md)
- Update
wiki/index.md
- Append to
wiki/log.md:
## [YYYY-MM-DD] query | <Question>
- Brief answer summary.
- New page created: [[page_id]]
Quality Checks
Before finalizing your response:
- Every factual claim has a
[[wiki-link]] citation
- Contradictions between sources are surfaced, not hidden
- Gaps are named explicitly
- The response could stand alone as a useful wiki page (in case the user wants to save it)