Literature Search

Literature Search

Goal

Find the smallest useful set of relevant scientific sources for the user's research question.

The job of this skill is not to dump papers. The job is to search the right sources, keep the query scoped, and return a candidate set that is useful for appraisal or synthesis.

This skill finds literature. It does not by itself prove a claim, judge full study quality, or synthesize a final evidence conclusion.

Default Posture

• source-aware search before broad web search
• narrow query before endless expansion
• quality of candidate set over quantity
• distinguish discovery from proof
• keep source trust visible

When To Use

Run this skill when:

• the user asks to find papers or studies
• a claim needs supporting or conflicting literature
• the next step is scientific discovery, not final judgment
• scientific-method-selector routes to literature search

This is a strong checkpoint before:

• study appraisal
• evidence synthesis
• claim-strength evaluation
• research-gap analysis

Scope Boundaries

In scope:

• define the research question in searchable terms
• choose the right scholarly sources
• generate a focused candidate set
• note obvious source limitations

Out of scope by default:

• full systematic review protocols
• full-text extraction from every source
• final evidence verdicts
• pretending search results alone settle the claim

Search Workflow

1. Restate the research question in one sentence.
2. Extract the key concepts:
   • population or domain
   • intervention, exposure, or mechanism
   • outcome
   • study type if relevant
3. Choose source scope:
   • PubMed / PMC for biomedical and health topics
   • arXiv for CS, math, physics, and preprint-heavy technical work
   • Semantic Scholar / OpenAlex for broader discovery
4. Build the narrowest useful query first.
5. Expand only if the first pass is too sparse.
6. Return a candidate set with source and trust notes.

Query Design Rules

• prefer explicit outcome terms over vague topic labels
• prefer study-type filters when quality matters
• avoid overbroad queries on the first pass
• separate core concepts from optional modifiers
• use synonym expansion only when recall is too weak

Source Heuristics

• prefer domain-native sources before broad aggregators
• treat PMC as full-text support for relevant biomedical items
• treat arXiv as discovery-rich but preprint-sensitive
• use broader scholarly search for cross-domain or exploratory work

If source availability is uneven, say so instead of hiding the blind spot.

Preprint Handling

When preprint-heavy sources such as arXiv are part of the search scope:

• mark the candidate set as mixed or preprint-heavy when appropriate
• keep preprints visible rather than blending them into peer-reviewed results
• prefer peer-reviewed corroboration when the user is asking for stronger support
• if no reviewed literature is found, say the search is useful for discovery but limited for strong claims

Escalation Conditions

Pause and keep the search narrow when:

• the research question is still vague
• multiple distinct claims are bundled into one search
• the desired scope is too large for one pass
• the source choice itself materially changes trust or coverage

If those conditions persist, recommend scientific-method-selector, intent-clarifier, or a narrower claim before expanding the search.

Output Contract

Always return:

1. research_question
2. search_scope
3. query_logic
4. candidate_sources
5. peer_review_status_notes
6. preprint_load (none, low, mixed, high)
7. coverage_limits
8. recommended_followup
9. next_step

Guardrails

• Do not treat search hits as confirmed evidence.
• Do not hide whether sources are preprints, peer-reviewed, or mixed.
• Do not expand the query so much that relevance collapses.
• Prefer a smaller, more relevant candidate set over paper dumping.
• If the question is too broad, narrow it before searching wider.
• If preprints dominate the candidate set, warn that claim strength remains provisional.

Anti-Patterns

• paper dumping
• one giant query with no structure
• mixing unrelated claims into one search
• broad search with no trust notes
• assuming first-page results reflect the whole evidence base

Example

User request:

Find research on whether time-restricted eating improves insulin sensitivity in adults.

Expected shape of response:

1. research_question: whether time-restricted eating improves insulin sensitivity in adults
2. search_scope: PubMed, PMC, broader scholarly discovery only if needed
3. query_logic: adults + time-restricted eating / time-restricted feeding + insulin sensitivity / insulin resistance + human study filters
4. candidate_sources: focused list of relevant studies and reviews
5. peer_review_status_notes: PubMed-indexed studies likely peer-reviewed; broader sources may include preprints and should be flagged
6. coverage_limits: outcome terms vary; related metabolic endpoints may need separate pass
7. recommended_followup: study-quality-appraiser
8. next_step: appraise the strongest candidate studies before making a claim-strength judgment