Systematic Literature Review with ChatGPT, Claude, or Gemini: A Sources-First Workflow

A chatbot can save you real time on a literature review. It can sharpen a vague topic into an answerable question, draft Boolean search strings, build a screening plan, and compare dozens of abstracts faster than you can read them. What it cannot do is replace the search itself. When you ask a chatbot to recall "the ten most important studies" on a topic, it answers from patterns in its training, not from a database, and the result can include papers that look real but are not. This guide shows a sources-first workflow that keeps the speed and removes that risk, built around the Systematic Literature Review Assistant (AI Prompt That Won't Invent Sources) prompt.

The short answer

Use AI in two separate stages. First, let it plan the review without naming any studies: question, databases, search strings, criteria, screening stages, and an extraction template. Then run the searches yourself in real databases. Finally, paste the papers you found and let AI synthesize only those papers, citing each one by an ID you assigned. Every claim then traces back to a source you actually have open.

Why the order matters

Language models are good at writing text that sounds like a literature review. That is exactly why they are risky as a search engine. A model asked to list studies will produce plausible authors, years, and titles because that is what a reference list looks like. A one line reminder to "verify all citations" does not fix this, because the problem is in the request, not the follow up. The fix is structural: never ask the model to supply sources, and require it to cite only from text you pasted.

How the two mode prompt works

The upgraded PromptDig prompt is split into two clearly labeled modes.

Mode 1: Plan the review (no papers yet). You give a research question, discipline and review type, scope, focus areas, and citation style. The model returns:

1. A refined question in a PICO, PEO, or SPIDER table, whichever fits your field.
2. Suggested databases for your discipline and why each one matters.
3. Boolean search strings for two databases, with synonyms, truncation, and field tags.
4. Inclusion and exclusion criteria tied to your scope.
5. A screening plan in PRISMA 2020 stages, with blank counts to fill in.
6. A data extraction template with a "not reported" column.
7. Candidate themes and frameworks, labeled as hypotheses rather than findings.

The hard rule in Mode 1 is simple: no studies, authors, years, titles, journals, DOIs, or citations. If you ask for studies, the prompt tells the model to reply that you should run the search and paste the results into Mode 2.

Mode 2: Synthesize the papers you paste. You paste abstracts, full text excerpts, or database export rows, each tagged S1, S2, S3 and so on. The model returns an evidence table, four to six themes, where the sources agree and disagree, methodology strengths and weaknesses, research gaps, a narrative synthesis, and a source manifest that maps each ID to its exact title and the claims that rely on it. Citations appear only as [S#]. Anything the pasted papers do not support is marked NO SOURCE or left out.

A step by step workflow

Here is how a full systematic review can move through PromptDig prompts, one stage at a time.

Step 1: Turn your topic into an answerable question

Run Mode 1 with your draft question. If you want a dedicated table for this step, use the Systematic Review PICO Table from Question Notes prompt. A good question names the population, the exposure or intervention, and the outcomes you will measure.

Step 2: Build and test your search strings

Take the Boolean strings from Mode 1, or refine them with the Systematic Review Search String from PICO prompt. Then test them inside each database. Every database has its own syntax and field tags, so expect to adjust. Save the final strings and the date you ran them for your methods section.

Step 3: Run the searches and screen titles

Search your chosen databases, such as PubMed, Scopus, Web of Science, PsycINFO, ERIC, IEEE Xplore, or Google Scholar, depending on your field. Export the results with titles and abstracts. For a first pass, the Title-Only Systematic Review Screener prompt sorts pasted titles into include, exclude, or unclear. The guide on how to screen paper titles for a systematic review walks through that stage in detail. Final decisions stay with you and your second reviewer.

Step 4: Keep a screening log

Record counts at every stage: records identified, duplicates removed, records screened, full texts assessed, and studies included, with reasons for exclusion. The PRISMA Screening Log from Pasted Abstracts prompt helps you organize this from what you pasted.

Step 5: Extract data and synthesize with Mode 2

Assign each included paper a stable ID and paste them into Mode 2 in batches. For a structured extraction pass, the Elicit Evidence Table from Paper Inventory prompt builds a table from your paper list. Mode 2 then turns the set into themes, agreements, disagreements, and gaps, citing only your IDs.

Step 6: Report and clean your references

Write your flow diagram text with the PRISMA Flow Text from Screening Counts prompt, check your manuscript against reporting items with the PRISMA Systematic Review Checklist Assistant, and format the papers you actually used with the APA 7 Reference List Cleaner from Pasted Citations.

How to paste sources so the synthesis stays clean

• Tag every paper once. Use S1, S2, S3 and keep the same IDs in every follow up message. If you add papers later, continue the numbering instead of restarting.
• Paste the exact title with each abstract. The source manifest repeats the title as pasted, which makes checking easy.
• Work in batches. A batch of ten to twenty abstracts is easier to review than a hundred at once. Ask for a combined synthesis after the batches.
• Say when you only have an abstract. Abstracts often leave out sample details or limitations, and the prompt is told to write "not reported" rather than guess.
• Keep your bibliographic details. Mode 2 builds formatted references only from details you pasted and flags missing fields.

What a good Mode 2 answer looks like

On the prompt page, the worked example uses four clearly labeled fictional placeholders, Example Paper A to D, on remote work and career growth. The evidence table shows design, sample, key finding, and limitation for each. One placeholder has an abstract that does not state its design, so the table says "not reported" and the synthesis gives it less weight. No pasted source tests gender differences, so any gender claim is marked NO SOURCE and listed as a gap. That is the behavior you want: honest about what the evidence does not cover.

Systematic, scoping, and narrative reviews

A systematic review follows a predefined protocol with an explicit search, criteria, screening, and reporting, often against PRISMA. A scoping review maps the size and nature of a literature, usually to find out what exists before a narrower question is asked. A narrative review is a more flexible expert overview without a fully reproducible search. The two mode prompt works for all three. Choose your review type in the inputs so the plan matches the level of rigor you need.

Common mistakes

• Asking the chatbot to "find" studies instead of planning a search.
• Mixing planning and synthesis in one conversation without clear source IDs.
• Treating suggested themes from Mode 1 as findings.
• Letting the model decide inclusion without a human check.
• Pasting a reference list from the model into your paper without opening each source.

FAQ

Can ChatGPT do a literature review for me?
It can plan your review and synthesize papers you supply. It should not be your search engine. Run the searches in real databases and paste what you find.

Why do chatbots sometimes produce references that do not exist?
Because they generate text that resembles a reference list instead of looking records up. The two mode design avoids this by never asking for sources and citing only what you pasted.

Can I use this for a PRISMA systematic review?
Yes, as an assistant. It drafts the plan, screening log structure, and synthesis, while screening and extraction decisions stay with the human reviewers.

Does this work in Claude, Gemini, or with uploaded PDFs?
Yes. The prompt is plain text and works in any major chatbot. If you upload PDFs, still tag each file with an S-ID so the manifest stays traceable.

Is this better than research discovery tools?
It does a different job. Discovery tools help you find papers, and this prompt helps you plan the review and synthesize the papers you found. Many researchers use both.

Do I still need to check every reference?
Yes. Open each source and confirm the title, authors, year, and DOI before it goes in your paper, and follow your institution's policy on AI use.

Start your review

Open the Systematic Literature Review Assistant (AI Prompt That Won't Invent Sources) prompt, run Mode 1 on your question today, and come back to Mode 2 once your searches are done. For more research prompts, Browse more prompts. If you have a review workflow that helps your lab, Share a prompt.