academic-literature-review

Academic literature review

Use this skill for one scholarly-review workflow: retrieval, evidence resolution, relevant graph expansion, and evidence-grounded synthesis. The MCP server supplies live scholarly metadata and supported full text; this skill supplies the order of operations and evidence boundaries.

Required input and output

Parse the research question and every explicit constraint before calling a tool. Preserve requested scope such as date range, language, study type, population, intervention, comparator, outcome, and identifier constraints. If a required constraint is missing or ambiguous, ask a focused follow-up question before proceeding.

Return a traceable review. For each material claim, identify the stable paper identifier returned by the tools and label the evidence as one of:

• Metadata-only
• Abstract-supported
• Full-text-supported
• Non-text evidence not interpreted

Disclose provider degradation, unavailable full text, identifier conflicts, and unsupported non-text evidence in the result. Keep retrieved rank separate from analytical importance.

Mandatory workflow

Follow these steps in order.

1. Parse the research question and explicit constraints.
2. Call search_papers for the baseline scholarly retrieval. Do not substitute ordinary web search for this baseline.
3. Inspect the search result’s provider-degradation information. Use provider_status only when the returned status is insufficient to explain material coverage loss, and disclose that loss.
4. Keep each returned retrieval rank distinct from any later judgment about analytical importance.
5. When a paper is central to a factual claim, call get_paper to resolve its identifiers and metadata.
6. Call get_full_text when the requested analysis depends on details that metadata or an abstract cannot safely support.
7. Call get_references or get_citations only when backward or forward chaining is relevant to the research question. Do not expand graph breadth without a research reason.
8. Distinguish explicitly between discovered metadata, abstract evidence, parsed full-text evidence, and model synthesis.
9. Cite claims to stable paper identifiers and state the evidence level beside the claim or in the claim’s evidence record.
10. Report unavailable full text, provider degradation, identifier conflicts, and unsupported non-text evidence explicitly.
11. Never fabricate a DOI, quotation, page number, sample size, effect estimate, or study result.
12. For a requested literature review, synthesize disagreements and limitations rather than collapsing all papers into a single consensus statement.