How I Produced a 45‑Page, SCI‑Ready Literature Review Using Claude Code

AI as a High‑Quality Information Processor

Large language models obey the "Garbage in, Garbage out" principle: reliable output requires supplying clean, relevant literature as input.

Preparing High‑Quality Sources

Use recent (last five years) open‑source databases such as arXiv, PubMed, and bioRxiv, together with closed‑access journals (IEEE, Springer, Nature, Elsevier). Import the PDFs into a personal Zotero library via the Zotero‑MCP tool; loading many PDFs directly into Claude Code would exceed context limits.

Example Literature Database

For coronary CTA deep‑learning segmentation, a Zotero collection of the relevant papers was created, enabling rapid summarisation through the MCP interface.

Using Claude Code in Plan Mode

Switch Claude Code to plan mode and feed one or two top‑journal review articles (e.g., Nature Reviews Cardiology , impact factor 44.2) so the model can learn the desired structure and style, producing a CLAUDE.md writing guide.

Generating the Review Plan

In plan mode, Claude Code accesses the Zotero collection and runs MCP searches on arXiv and PubMed with keywords such as "coronary artery segmentation". It updates CLAUDE.md while simultaneously creating an IMPLEMENT.md that outlines the review framework. The framework in IMPLEMENT.md must be manually inspected and edited as needed.

Iterative Updates and Supplementary Research

When new papers appear, update the two Markdown files. Additional reports from ChatGPT and Gemini's DeepResearch can be added to the knowledge base and fed back to Claude Code to refine the documents.

Automated Draft Generation

Claude Code writes the full literature‑review draft based on the finalized Markdown guides, requiring no further human intervention except a final review. The draft is converted to Word via Pandoc, yielding a 45‑page, 13,813‑word document with 113 references and a layout that follows Nature's style.

Final Polishing and Submission

A brief manual edit adds figures and adjusts the reference format to meet the target journal’s guidelines, preparing the manuscript for submission to a Q1 or Q2 journal.

Core Insight

When the literature set, writing framework, and style guidelines are firmly established, the remaining work is pure information processing and arrangement—tasks that AI performs efficiently, dramatically reducing manual effort.

Open‑Source Skill

The workflow is packaged as the medical-imaging-review skill, available at https://github.com/luwill/research-skills. Users need Zotero‑MCP or the Zotero API and a curated collection of papers for their specific review topic; the skill currently targets medical‑imaging AI but can be adapted to other domains.