Overview
Video not available yet
- Overview
-
Six public steps on using AI to build a corpus of academic articles for a literature review, from a set of seed articles to a deduplicated Zotero library with full texts. A review of post-socialist transformation in North Korea, China, and Vietnam is the worked example.
- Scope
Six steps on using an AI agent to build the article corpus for a literature review, from a set of seed articles to a deduplicated Zotero library with full texts, ready for analysis.
- Series
Part 1 of 3 · AI x Literature Review
- Format
Tutorial · 6 steps · about 2 hours to read
- Worked example
Post-Socialist Transformation (North Korea, China, Vietnam), a review run in October and November 2026
- Next
Literature Review II · From Corpus to Arguments analyses the corpus built here.
A literature review begins with a corpus, and the corpus decides what the review can find. Searching by hand is slow and narrow; handing the search to an AI agent without supervision is fast and arbitrary. This tutorial shows how to let the agent do the searching, reading, scoring and filing while the researcher keeps the decisions that define the review.
The workflow has six steps. You set up a Zotero library and a workspace the agent can use (Step 0), collect the seed articles that define the review (Step 1), derive search terms and choose the journals to search (Step 2), search and score the results against the seeds (Step 3), extract and score the works the seeds and the best candidates cite (Step 4), and select the corpus, collect the PDFs and export it (Step 5). At every step the agent proposes and the researcher approves.
Every page follows one review, on post-socialist transformation in North Korea, China, and Vietnam. Its 264 seed items led to 3,060 searches, 19,186 scored search results, 3,684 scored cited works and a selection of 1,038 works. The Zotero library exported for analysis on 27 October 2026 held 1,261 works, 434 of them chosen for full reading and 309 with a full text. That corpus is the one Literature Review II analyses.
What you need
- Zotero, the desktop application and the browser Connector, and a free Zotero account that can own or join a group library.
- An AI agent that can work with files, such as a desktop coding agent (Claude Code, Codex), so that it can read the seed PDFs, run searches, write results to the workspace and call the Zotero web API. A chat interface is enough to follow the reasoning but not to run the pipeline.
- Library access through your university, for the publisher sites and the Chinese and Korean databases. The worked example used the NCCU library and its proxy.
- A set of seed articles on your topic, as Zotero items with PDFs where you have them. Step 1 explains what makes a good seed set.
How to read this
Each step page states the decision the researcher makes, what the agent does, and the files the step leaves behind. Tables give the real counts from the worked example, and screenshots show the Zotero library as it was during the review. The two example pages imitate working pages from the review, with their links made inactive. Where a number could not be checked against the review's own records, the page leaves it out.
Content
Part 1 of the AI x Literature Review series. Read the steps in order. Each step gives the decision the researcher makes, the work an AI agent does, and the real counts from the worked example. The corpus built here is the one that Literature Review II analyses.
- 0Set upCreate a Zotero group library with numbered collections, a review workspace in which every step keeps its own folder, and an AI agent that can read and write both. Fix the design of the review before anything is searched.
- 1SeedsCollect the articles that define the review, 264 items in four seed collections, add the researcher's own drafts as reference documents, and turn them into a weighted comparison corpus of 121 records.
- 2Search prepDerive search terms from the seeds, from 49,794 candidates to 10 general terms and three country sets, choose the journals for each search, and translate the terms for the Chinese and Korean databases.
- 3Search and scoreRun 3,060 queries across journals and open indexes, pool and deduplicate the results, and score 19,186 candidates for their similarity to the seeds.
- 4Extract and scoreExtract the 4,005 references cited by the seeds and the strongest candidates, validate them against Crossref, and score the 3,684 cited works on the same scale.
- 5Zotero, PDFs and exportSelect 1,038 works by four routes, import them into Zotero, choose 300 for full reading, collect the PDFs through the library, merge duplicates and export the corpus for Literature Review II.
Workshops
Resources
- Relevant Literature
-
Browse references and supporting literature for this training module in the Organization of Power Lab Zotero library.
Open the lab library on Zotero
Worked example from a literature review on post-socialist transformation run by the researcher, Organization of Power Lab, National Chengchi University (NCCU), October and November 2026. Search, scoring and filing were carried out by an AI agent working from the lab's literature-review skills, with every cutoff and selection approved by the researcher.