paper-ai-search

name	Paper AI Search
description	Use when the user describes a research need in natural language and wants AI-assisted query expansion before searching for papers, including requests to find recent work on a topic, related papers in a direction, or similar Chinese requests for AI paper search.
allowed-tools	["searchArticles","readPaper"]

Paper AI Search

Use this skill when the user gives a research question, a direction, or a vague paper-finding intent rather than a ready-made keyword list. The core workflow is query expansion first, then cross-source retrieval.

Workflow

1. Rewrite the user's natural-language request into a cleaner academic search intent.
2. Extract 3-6 compact keywords, each usually 1-3 words, favoring terms that would plausibly appear in titles or abstracts.
3. Produce one refined semantic query sentence that captures the intent at a higher level.
4. Search across all relevant academic sources with the expanded keywords rather than staying on a single source.
5. Show the expanded keywords back to the user so the search process is inspectable and easy to refine.

Query Expansion Rules

• Cover the central technical concepts rather than relying on generic words.
• Mix narrower technical terms with slightly broader related terms to balance precision and recall.
• Avoid low-information tokens such as method, model, or improvement unless they are part of a recognized phrase.
• If the request is underspecified, infer and separate the task, method family, and application setting.

Retrieval Sources

Search these sources by default:

• arxiv
• huggingface
• semantic-scholar
• biorxiv
• pubmed
• pubchem

Output Structure

• Briefly restate the actual search intent.
• List the Expanded keywords explicitly.
• Include a Semantic query line when it adds clarity.
• Present the ranked paper shortlist. Each entry should include:
  • original English title
  • authors
  • date
  • source
  • one sentence on why it matches the user's intent
• End with Next step guidance:
  • narrow the scope
  • branch into a promising subtopic
  • or move straight to summary / deep reading

Quality Rules

• Do not present query expansion as ground truth. It is only a retrieval strategy.
• Do not pad the answer with papers that are visibly unrelated to the expanded query.
• If the results are diffuse, state that the user request is still too broad.
• Preserve technical terms in their standard English form when that improves precision.