| name | paper-wiki-query |
| description | Answer questions using the persistent paper literature wiki. Use when Codex needs to query the paper wiki, answer based on the wiki, explain what the wiki knows, summarize a topic, compare papers or claims, retrieve wiki pages, synthesize an answer with page citations, decide whether an answer is worth filing, or revise WikiQuery behavior. |
Paper Wiki Query
Query answers a user question using retrieved wiki pages. It builds on paper-wiki-search.
This skill is schema-bound. Before changing query behavior, read references/query-schema.md.
Examples
- User: "what does the wiki know about diffusion model evaluation?" -> use
paper-wiki-query.
- User: "compare these two claims based on the wiki" -> use
paper-wiki-query.
- User: "find pages mentioning MATH benchmark" -> use
paper-wiki-search, not query.
Workflow
- Use
paper-wiki-search to retrieve primary and expanded pages.
- Load the retrieved page markdown.
- Ask the model to synthesize an answer using only those pages.
- Cite wiki pages with
[[page_key]].
- Decide whether the answer is a good answer worth filing.
- File only when the answer passes the good-answer bar and
fileAnswer is true, or when fileAnswer is auto and model output says shouldFile.
Query Depth
For query mode guidance, also read paper-wiki/references/query-modes.md.
Filing Answers
Good answers can compound into the wiki, but only through the explicit filing path.
A good answer is durable, source-backed, and reusable. It should synthesize multiple relevant wiki pages, cite them clearly, and add a comparison, conclusion, tension, gap, or reusable framing that is not already captured by an existing page.
- If
fileAnswer is false or absent, return the answer without writing a page.
- If
fileAnswer is auto, file only when the model marks shouldFile true and the answer meets the good-answer bar.
- If
fileAnswer is true, still file only good answers; otherwise return the answer and explain why it was not filed.
- Do not file narrow lookups, weakly supported answers, or answers with insufficient retrieved evidence.
Boundary
Do:
- Synthesize across pages instead of listing summaries.
- Return answer markdown and citations.
- Build a candidate synthesis page object when useful.
- Keep the retrieval result inspectable so the caller can see which pages informed the answer.
Do not:
- Answer paper-specific questions from general model knowledge when retrieved pages are insufficient.
- Write every answer automatically.
- Create overview pages.
- Bypass retrieval.
Context Discipline
- Quick questions should not force a deep read.
- Standard questions should use the retrieved primary pages and only the linked pages needed for context.
- Deep questions may read more broadly, but still need citations for non-obvious claims.