Do a literature review with an agent that finds, screens and summarises real papers
Ground the agent in real paper records before it writes anything: a Zotero library, a folder of PDFs, or a scholarly search tool such as a paper search MCP. Have it screen papers against inclusion criteria you wrote, with a reason for each decision, and extract findings into a table that names the page or section. Then check that every reference resolves to a real DOI or arXiv ID. Studies of chatbot-written reviews found many fabricated or wrong references, so the reference check is not optional.
You get a screening log, an evidence table with one row per included paper, a synthesis that cites only papers you hold, and a reference list where every entry resolves to a real record.
- Difficulty
- Intermediate
- Time
- Half a day to a few days, depending on the number of papers
- Sources
- 5 · 5 independent
- Updated
- 2026-10-13
- Cost:
- Included in a Claude, ChatGPT or Gemini plan. Zotero and the scholarly APIs used here are free.
The prompt
Fill the blanks here, or leave them for your agent to ask. Pick the tab for your agent. The hand-off button gives any agent this page as markdown.
You are helping me with a literature review. Work only from papers you can actually open. Never cite a paper from memory. Question: [research question] Inclusion criteria: [population or system, methods, years 2019-2026, peer-reviewed or preprint, language] Exclusion criteria: [e.g. opinion pieces, studies without evaluation] Sources: [my Zotero collection "X" / the PDFs in ./papers / search with the paper search tool in arXiv, Semantic Scholar and OpenAlex] Step 1, search: run 3-5 searches with the terms I approve: [terms]. Log each query, the source and the number of results. Step 2, screen: for each candidate, decide include, exclude or unsure from the title and abstract, and give a one-line reason tied to a criterion. Work in batches of 20 and show me the first batch before you continue. Step 3, extract: for each included paper, fill a table with: citation, DOI or arXiv ID, design, sample or dataset, main finding, limitation, and the section or page where you found the finding. Step 4, synthesise: write 800 words grouped by theme. Cite only papers in the table. Mark where findings conflict and where evidence is thin. Step 5, verify: for every reference, check that the DOI or arXiv ID resolves and the title and authors match. List any that do not. Do not invent page numbers, sample sizes or results. If a detail is not in the paper text you read, leave it blank. Do not submit, email or upload anything.
Human checkpoints: the agent stops and asks you
- Write and approve the research question and inclusion criteria before the agent screens anything.
- Check a sample of include and exclude decisions against the papers.
- Confirm every reference resolves and read the papers behind your main claims before you submit or publish.
Fabricated or misdescribed references can end up in a thesis, grant or paper under your name. You are responsible for every citation.
What you need
Free reference manager. The local API must be enabled for local MCP connectors.
- A Zotero or paper search MCP servermcp
Community projects. Read the code or use read-only keys.
- Access to full textsaccount
Your library login for paywalled papers. Agents cannot read paywalled text on their own.
- Scholarly metadata APIs for the reference checkdata
Crossref, OpenAlex, Semantic Scholar or the arXiv API. arXiv asks for no more than one request every three seconds.
Steps
Write the question and criteria
Write the research question and include and exclude rules in a file such as criteria.md. The agent screens against this, so make each rule checkable.
Check: Someone else could screen a paper with your criteria and reach the same decision.
Ground the agent in real records
Connect Zotero, put PDFs in a folder, or use a paper search tool. Do not ask a chat model to list papers from memory.
Check: Every candidate has a DOI, arXiv ID or a PDF you hold.
Run and log the searches
Approve the search terms. Have the agent log each query, source and hit count so you can report your method.
Check: You have a search log with dates and counts.
Screen in batches with reasons
Review the first batch closely. Correct the agent's reading of your criteria, then let it continue. Spot-check excludes as well as includes.
Check: You agree with at least the sample of decisions you checked, and each decision names a criterion.
Extract findings with page references
Have the agent fill the evidence table from the full text, with the page or section for each finding.
Check: Opening a random row's paper at the given page shows the stated finding.
Synthesise from the table only
Ask for a synthesis by theme that cites only rows in the evidence table and marks conflicts and gaps.
Check: Every citation in the synthesis maps to a row in the table.
Verify every reference
Resolve each DOI or arXiv ID and compare title, authors, venue and year. Fix wrong details and remove anything that does not resolve.
Check: All references resolve and match. None were added from memory.
Read the papers behind your main claims
Read the key papers yourself before you submit, publish or send the review.
Check: You can defend each main claim from the paper itself.
What goes wrong
The agent cites papers that do not exist or mixes real titles with wrong authors.
Fix: Ground it in records you hold and check every DOI or arXiv ID against a metadata API.
Read the lesson →A real paper is cited for a finding it does not report.
Fix: Require the page or section for each finding and spot-check rows against the PDF.
Read the lesson →Asked to summarise a citation it cannot find, the model writes a plausible study anyway.
Fix: Tell it to answer "not found" and only summarise papers you have given it.
Paywalled papers are skipped, so the review leans on open access work.
Fix: Download paywalled papers through your library and add them to the folder or Zotero.
Fields such as sample size or effect size get filled in when the paper does not report them.
Fix: Tell the agent to leave unreported fields blank, and check a sample of rows.
The paper search API rate-limits the agent.
Fix: Slow the requests down. arXiv's terms ask for one request every three seconds on one connection.
Evidence
Two or more public accounts from different people or sites describe the same job. We summarise each source in our own words. Open the original before you rely on it.
- Blog2026-04-23· Claude CodeAcademic use cases with Claude
A hydrology researcher uses Claude Code with a paper search MCP and a Zotero MCP to find papers and write notes from his own annotations. He finds it faster than browsing Google Scholar and has not yet tested a full written review.
Pin Shuai · groundwater.usu.edu - Blog2026-06-27· Claude CoworkConnect your local Zotero library with Claude Cowork
The guide connects a local Zotero library to Claude through a community connector that reads full text and annotations. It warns that setup takes long for large libraries and that the default search model can give poor results.
effortlessacademic.com - Web2025-04-24· ChatGPTUsing ChatGPT to review the literature, a cautionary tale
Surgeons had ChatGPT-4o write a literature review. Of 21 citations, 5 were correct, 8 had errors and 8 could not be found in PubMed. Asked about a fake citation, it invented a study.
Kate Manley and others · pmc.ncbi.nlm.nih.gov - Web2023-09-07· ChatGPTFabrication and errors in the bibliographic citations generated by ChatGPT
Across 84 short literature reviews, the authors found 55 percent of GPT-3.5 citations and 18 percent of GPT-4 citations were fabricated, and many real citations had wrong details.
William H. Walters and Esther Isabelle Wilder · pmc.ncbi.nlm.nih.gov - Blog2026-01-21Hallucinated citations in NeurIPS 2025 accepted papers
GPTZero says it found 100 hallucinated citations in more than 50 accepted NeurIPS papers, including invented authors, fake arXiv IDs and real titles with the wrong authors, each confirmed by a person.
GPTZero (Nazar Shmatko and others) · gptzero.me
- Scite tutorial, systematic review with ChatGPT and the Scite MCPchecked 2026-10-13
- arXiv API terms of use (rate limits)checked 2026-10-13
- OpenAlex documentationchecked 2026-10-13
- Semantic Scholar APIchecked 2026-10-13
Report how it went
Every report makes this page better for the next person and the next agent. Reports are reviewed before they count. Agents can report with the report_run tool.
Loading reports…