nature-academic-search

A router for coordinated, multi-step academic literature workflows that go beyond a single MCP call. It covers multi-source literature search, citation verification, MeSH search-strategy building, citation-file management for .nbib/.ris/.bib conversion, and reference management such as BibTeX generation, related-article discovery, and ID conversion, working through MCP tools including PubMed, CrossRef, arXiv, Scopus, and ScienceDirect. It is designed to trigger on general literature needs during academic writing even without the word "Nature", and recognizes Chinese phrasings for searching, reviewing, and managing references.

The skill is split into two layers. A static layer holds versioned, reusable fragments such as the MCP tool inventory and shared modules plus source routing and operational rules. A dynamic layer, this router file plus manifest.yaml, detects which workflow the user needs and loads only that workflow, reaching for shared modules and scripts on demand. The design deliberately avoids applying search logic from memory; fragments are always loaded from disk.

A five-step routing protocol governs each invocation. It first reads manifest.yaml and the always-load core files for the tool inventory and the T1 to T2 to T3 source-routing guide. It then maps the user's need to one or more workflow values, states the detected workflow in a short line, and loads the matching workflow fragments under references/workflows without reading every workflow. It runs the workflow by applying core tools and routing first, then the workflow-specific steps, then shared modules and scripts such as deduplication, citation parsing, search strategy, and RIS/BibTeX formatting only when a step needs them. It reads deeper references and scripts, such as the full source-tier classification or the no-MCP fallback discovery search, strictly on demand. Failures of individual tools are reported while continuing with the rest, terms are broadened when there are no results, and manual generation from fetched metadata is a fallback when a script fails twice. This structure mirrors the other nature-* skills.