How Graduate Students Can Use AI to Help Complete a Literature Review: My Research Writing Tools and Practical Experience

How Graduate Students Can Use AI to Help Complete a Review: My Research Writing Tools and Practical Experience

Matrix is SSPAI's writing community, where we advocate sharing genuine product experiences and practically valuable lessons and reflections. From time to time, we select the highest-quality articles from Matrix to showcase the most authentic experiences and perspectives of users.

The article represents the author's personal views, and SSPAI has made only minor modifications to the title and formatting.

Abstract: From determining a topic and searching the literature to organizing research evidence, writing the main text, drawing mechanistic diagrams, and checking formatting before submission, completing a review involves a great deal of repetitive and tedious work. Based on a review I participated in completing that has been accepted by a journal, this article shares how to use AI, Zotero, Keynote, and AI Agents to assist in research writing, and demonstrates the actual operations, usage experience, and areas requiring manual verification at each stage. The point is not to let AI write papers for us, but to explore how to use digital tools to reduce repetitive labor while retaining the necessary scientific judgment.

Image sources: The cover image is from Beku Kanomi on Unsplash; the images in the body mainly come from my personal Keynote creations and software screenshots.

If you are currently a graduate student facing the hard requirement that "you must publish a small paper to graduate," yet you are in a period of confusion with no guidance and no direction, full of motivation but not knowing where to start—this article may be able to help you.

I have been through this process myself. Later, I began trying to introduce AI tools into different stages of review writing, and combined Zotero, Keynote (macOS) or PowerPoint (Windows), and WorkBuddy to form my own research writing workflow.

In this article, I want to fully break down this battle-tested method for you.

My Results

To make this workflow easier to understand, I want to use a review paper I completed not long ago as an example to explain how it was advanced in the actual writing process.

This paper was ultimately accepted by Food Science and published online first. As the first author, I completed everything from topic selection and literature organization to main text writing and mechanistic diagram drawing on my own. I am grateful to my teachers and the reviewers for their suggestions, and I also revised it patiently and carefully.

At different stages, I tried using AI and digital tools to assist in completing part of the work, including:

• Topic selection stage: Using Gemini to help me explore research directions, compare topics, and organize the article framework.

• Literature stage: Using Zotero to collect and manage literature, and using AI to assist in organizing abstracts.

• Writing stage: Using Qwen to summarize the research materials already organized, assisting in sorting out chapter logic and textual expression.

• Diagramming stage: Using ChatGPT to obtain preliminary composition references, then making formal figures with Keynote.

• Submission stage: Using WorkBuddy to assist in checking document formatting and references.

It should be noted that, as an engineering student in the food field, this article is also a review in the engineering field. The above tools are mainly used to clarify my thinking and direction in each key work stage of writing an academic article, enabling me to get started better and lowering the difficulty. They may not suit your field and may require some appropriate adjustments and personal judgment. But I think the direction and thinking embodied here are very valuable.

So, how exactly should one begin writing a qualified, high-quality review?

Preparation Before Writing a Review: Focus and Literature Tiering

Reviews are highly "time-sensitive," mainly reporting recent research progress and gaps in a certain field regarding a "small pain point or small hot topic" (for example: the combined action mechanism of a specific substance or sensory quality improvement). Therefore, when choosing a review topic, avoid being broad at all costs; only a high degree of focus can thoroughly explain a small issue.

In literature selection, although I value cutting-edge results from recent years, I do not judge solely by year. A more scientific approach is to manage literature by tier according to function:

• Classic literature: Explains the research background, basic theories, and foundational methods (before citing, check whether there are subsequent errata or revised conclusions).

• Representative studies: Explains the evolution of key technologies or academic viewpoints.

• Frontier research: Reflects new breakthroughs and new applications in the past 3–5 years.

• Controversial studies: Horizontally compares differences in conclusions among different teams and limitations of experimental conditions.

Having clarified the above principles, let us now formally enter the topic: how can we make good use of AI tools to quickly write a review article?

Step 1: Use "Research Prompts" to Determine the Topic and Outline

Before starting to write a review, we need to first complete basic accumulation by reading recent reviews in the same field and figure out the industry's pain points. On this basis, rather than staring at a blank document in deep thought, it is better to use AI tools with built-in internet access (such as Gemini, ChatGPT, or domestically directly accessible Qwen) to help brainstorm ideas.

Asking AI questions requires polishing "research prompts." Below is the prompt template I use to determine review topics and excavate research gaps (Gemini is recommended):

        # Role Setting
You are a senior research expert and journal reviewer in the field of [fill in the discipline, e.g., Food Science and Engineering], and I am a graduate student in this field preparing an academic review.

# Task Objective
I plan to write a [Chinese / English] review article related to [fill in the general direction, e.g., high-value utilization of sea buckthorn pomace and quality improvement of flour products]. Please combine the domestic and international research status of the past 5 years (2021–2026) to help me excavate the research pain points or review gaps that still exist in this direction, and provide 5 specific topics with publication potential.

# Output Requirements
Please output each topic according to the following structure:
1. Proposed title (must be specifically focused, avoid being broad)
2. Core entry point and necessity of the review (what industry/academic pain point it solves)
3. 3–4 core discussion dimensions covered

# Research Rigor Constraints (must be strictly observed)
- It is strictly forbidden to fabricate papers, authors, journals, DOIs, or specific experimental values.
- Clearly distinguish between "consensus already confirmed by the academic community" and "inferences you make based on logic."
- For judgments lacking sufficient evidence, cautious academic expressions such as "may," "suggest," and "remains unclear" must be used.
- If viewpoints with academic controversy are involved, please also list the research perspectives and condition differences of both the pro and con sides.
      

After obtaining several topics, choose one that you think is not bad and whose logic you can roughly sort out, and begin the next step:

        # Task Objective
I have ultimately selected "【fill in the confirmed review title】" as the review topic. Please design for me a detailed three-level outline and table of contents that conforms to the standards of mainstream core journals (such as *Food Science* or SCI journals of the same level).

# Specific Requirements
1. Logical progression: Chapters must reflect a progressive relationship of "background pain points -> comparison of core mechanisms/treatment methods -> quality/application impact -> existing bottlenecks and prospects," and simple literature listing is rejected.
2. Writing guidance: Below each second-level/third-level heading, briefly note in parentheses "which research indicators should be mainly compared" or "what core scientific questions need to be answered" in that subsection.
3. Search term suggestions: At the end of the outline, extract the core Chinese and English search keywords needed to support each chapter.
      

Then, based on the knowledge base you already have, use multiple AI tools to verify whether this table of contents still has places that can be supplemented, improved, or revised, so that the logical structure of the entire review article becomes clearer.

Because my personal research direction is "sea buckthorn pomace," my review topic was also fixed on "solving the industry pain points of sea buckthorn pomace's taste and degree of utilization." I gave my needs to Gemini:

My larger project needs to involve the taste and product flavor of flour products, and I have some accumulation and understanding of Topic 1, so I ultimately chose Topic 1.

Then, according to Topic 1, I generated and improved the outline framework and determined the search terms:

After completing the topic selection and outline generation, we can move on to the next stage.

Step 2: Database Literature Search

After determining the topic and outline, I do not directly ask AI to search the literature for me. Instead, I first use it to break the research topic down into three dimensions: "core object, treatment technology, application direction," expand synonyms and generate search formulas adapted to the syntax of different databases, and then combine the synonyms of these concepts to form an executable search strategy.

But note: AI can help us expand keywords, but the final search results still require us to return to the database to search and screen.

Title: How Graduate Students Can Use AI to Complete a Review: My Research Writing Tools and Practical Experience

Body:

        # Task Objective
I am preparing to write a review titled "【Insert review title】". Please break down my research topic into three dimensions: "core research subject", "key processing technologies/methods", and "research results/application directions", and construct advanced search queries that I can copy directly into research databases.

# Specific Requirements
1. Keyword expansion: For each of the three dimensions above, list core terms in both Chinese and English, synonyms, near-synonyms, and common academic abbreviations.
2. CNKI search query: Provide 2–3 groups of search term combinations adapted to the CNKI advanced search syntax (distinguishing the "topic" and "title" fields, using SU=, TI=, and the AND/OR logical operators).
3. Web of Science (WoS) search query: Provide 2–3 groups of English Boolean search queries adapted to the WoS Core Collection (using the TS= topic field, quotation marks for exact matching, and the * wildcard, e.g., "sea buckthorn pomace" AND ("solid-state fermentation" OR pretreatment)).
      

Following the steps above, search results related to our topic will appear. For the first round of literature collection, I suggest judging by the title whether each paper is relevant to your review content: keep the relevant ones in Zotero and skip the irrelevant ones.

Step 3: Preliminary Screening of References

With the above preparation done, we come to the main event of writing a review—the extraction, screening, and management of abstracts.

Traditional review writing requires batch downloading, categorized management, and then reading a large number of papers... It is quite troublesome, and it looks so complicated that you really don't want to start. Let me share a good approach.

Using "Literature Abstracts" as the Core Input

Writing a review requires referencing the content of a huge number of papers, including but not limited to research background, experimental methods, core data, and conclusions. If you throw dozens of complete PDFs directly at AI, you will not only easily exceed the context limit, but also trigger serious "AI hallucinations" due to too much irrelevant detail. In fact, the highly condensed "literature abstract", written by the original authors, already contains the core elements above. Feeding abstracts instead of full texts to AI can greatly increase the number of papers processed in a single conversation, and significantly improve the accuracy and focus of information extraction.

Preparing Abstracts

After importing the preliminarily screened literature from CNKI or Web of Science into Zotero, there is no need to manually copy abstracts one by one; you can export them in bulk with the following 4 steps::

• In Zotero, select and select all the literature you want to use.

• Right-click -> select "Generate Report from Items"; Zotero will pop up a window that neatly lists titles, authors, abstracts, and other information.

• Select all and copy (Cmd + A, Cmd + C) the content of this report.

• Paste (Cmd + V) it into the AI tool's dialog box.

This is the preparatory work before feeding abstracts to AI. With such a literature abstract file, we can proceed with the following steps.

Screening Abstracts

After obtaining the abstract report, do not directly ask AI to generate the main text. I usually first feed it to Qwen for relevance grading and preliminary information extraction, removing irrelevant literature:

        # Task Objective
Please perform preliminary screening and core information extraction on this batch of literature based on the Zotero literature export report I provide.
Note: This task is limited to preliminary literature screening and abstract extraction. Please do not directly write the main text of the review.

# My Review Topic and Outline Direction
【Insert your review title or the current chapter topic】

# Output Format Requirements
Please first summarize the following 6 basic items in a Markdown table:
| Original literature number | Author and year | Research subject and core question | Key research methods | Main results and conclusions of the abstract report | Relevance rating (high/medium/low) |

Below the table, please list separately for "high-relevance" literature:
- Suggested chapter assignment: Which part of my outline is this literature suitable for?
- Blind spots to be verified: Key experimental conditions or data not provided in the abstract that must be confirmed by reading the full text.

# Rigor Red Lines
1. The literature numbers in the original report must be strictly preserved; do not shuffle the order.
2. It is strictly forbidden to fabricate experimental results based on paper titles, and strictly forbidden to add data that does not appear in the abstracts.
3. When conclusions differ across studies, record them truthfully and separately; wherever the abstract is vague, mark it uniformly as "[needs full-text verification]".

# Zotero literature report to be processed:
【Paste the report content generated by Zotero here (it is recommended to feed it in batches by subsection, about 15–20 papers at a time, to avoid the model missing details due to excessively long input)】
      

After completing the above steps, we obtain an abstract report that fits our topic very well, laying a solid foundation for writing the main text later.

Step 4: Writing the Main Text

A major taboo in review writing is rewriting and simply compiling literature abstracts one by one in order of publication date or research subject. A more scientific approach is to take "research questions or mechanisms of action" as the main thread—for example, horizontally comparing the effects of different treatment methods on the same raw material, or analyzing the differences in conclusions produced under the same experimental method due to different conditions and indicators.

When using AI to refine the first draft of a review chapter, I often use the following prompt:

        # Role Setting
You are now a senior scholar in 【insert discipline field】. I am writing a review paper "【insert review title】", and I now need you to help me write the specified chapter in the outline based on the literature abstract library I provide.

# Input Materials
1. Full outline background:
【Paste the complete outline to help AI understand the context boundaries】

2. The target chapter to be written this time:
【Insert the specific subsection title, e.g., 2.2 Effects of ultrasound and superheated steam pretreatment on the structure of dietary fiber from fruit pomace】

3. Zotero literature abstract report supporting this chapter:
【Paste the screened highly relevant literature abstracts】

# Writing Task and Core Requirements
1. Synthesize first, then discuss (refuse to pile up literature): Do not rewrite abstracts one by one in the流水账 style of "A studied what, B studied what". You must take "research questions or mechanisms of action" as the main thread, horizontally compare, classify, and connect the methods, experimental conditions, and conclusions of different papers, and point out common patterns and reasons for differences.
2. First draft writing: Write an academic first draft of this subsection, with a length of about 【e.g., 1000】 words. The language must be rigorous, objective, and plain, and conform to the academic norms of Chinese core journals.
3. Source annotation: Wherever viewpoints or data are cited in the text, the source must be strictly annotated at the end of the sentence in the format 【[original literature number, first author, year]】, so that I can trace and verify it later.
4. Factual boundaries: All statements must come 100% from the literature abstracts I submitted this time. It is strictly forbidden to use your own external pre-trained knowledge to fabricate data or conclusions out of thin air. If the existing abstracts are insufficient to support a certain argument, please mark it directly in the text with 【[lack of literature support here, additional search needed]】.
      

With the above prompt, each subsection of the outline only needs one batch of abstracts to build a first draft (I usually first use Gemini and Qwen to refine drafts for each chapter, then hand them to ChatGPT to supplement and polish against the core literature).

After the first draft is completed, I personally check the review through multiple rounds:

• Adjust some language expressions to better fit the style of academic articles.

• For expressions involving "data", I always go back to the references to verify whether the numbers are accurate.

• For expressions involving "proper nouns", I go back to the references to verify whether the names are correct and how to express them in English.

• For descriptions involving processes such as pretreatment, solid-state fermentation, and product conversion, I carefully read the literature, take notes, and then use AI tools to organize the core methods and processes in the literature for cross-checking. Wherever there are conflicts or inconsistencies, I compare them carefully to confirm the accuracy of my content.

This is work we must do ourselves. The subsequent figure-making process is directly affected by the accuracy of the review content.

Step 5: Using AI Tools and Keynote to Make Figures

A complete and qualified review article not only needs a complete structure and scientifically rigorous expression in content, but also needs clear and readable tables and intuitive, standardized mechanism diagrams. How to draw a figure that has substance and is suitable for use in a review article is not an easy project.

This is actually the most headache-inducing stage for me when writing a review, because I simply didn't know how to start. It was only after asking Gemini that I gradually got some ideas.

I first asked ChatGPT to generate a sketch for me based on part of the review main text, with the following prompt:

Article title: How Graduate Students Can Use AI to Help Complete a Review: My Research Writing Tools and Practical Experience

Body:

        # Role Setup
You are now a professional scientific illustration designer, skilled at creating clear, standardized mechanism diagrams (Mechanism Diagram / Graphical Abstract) for high-level academic journals.

# Task Objective
Based on the review body paragraphs I provide, first help me outline a "visual layout plan" for the mechanism diagram, then call the drawing tool to generate a high-resolution scientific mechanism sketch, as a visual reference framework for me to later trace and lay out in Keynote.

# Review Body Content to Be Visualized
[Paste here the review body paragraphs you have already checked, such as: raw material pretreatment, solid-state fermentation process, or product conversion mechanism]

# Specific Drawing Requirements
1. Logic and zoning: First briefly explain in text how you plan to divide the zones (for example: use a "left-to-right" process layout, or a "top-bottom layered" macro-micro mechanism layout), extract 3–4 core stages, and use clear arrows to indicate the direction of material conversion or action.
2. Visual style: Use a flat, modern 2D scientific vector illustration style (referencing BioRender or mainstream SCI journal figure styles). Pure white background, simple and restrained color scheme (no more than 3 main colors; fresh, soft academic colors are recommended). Cluttered 3D rendering, sci-fi light effects, or overly realistic complex backgrounds are strictly prohibited.
3. Elements and labels: Use intuitive scientific icons, simplified molecular/cell structures, or schematic graphics of experimental equipment as much as possible, reduce large blocks of text, and keep only short English core terms as placeholder labels at key nodes (so I can later replace them in Keynote with accurate Chinese or standardized terminology).
4. Rigor constraints: All reaction steps, material elements, and arrow directions appearing in the image must be strictly based on the body content I provide. Do not fabricate reaction pathways or experimental equipment that are not mentioned.
      

The first generated image may differ from your instructions, goals, and review content in many places, and may even have obvious font or image errors. That's fine, because in this process we have already gained something very important—a template!

From something originally imagined, it has become a practical thing with a "handle." We can use this framework to draw the mechanism diagram. And with this image, we can check it against the review content: what doesn't match? What obvious errors are there?

After several rounds of adjustment and review, I will trace it out in Keynote.

Since I personally use a Mac, some niche tools that only work on Mac are inconvenient for me to recommend to everyone. I also personally think there is no need to deliberately learn a new piece of software just to draw paper mechanism diagrams.

I made all my review mechanism diagrams with Keynote (Windows users can use PowerPoint).

The main drawing process is to reproduce, point by point, the image ChatGPT generated for us as a reference. Some image assets need to be found; some cannot be found and need to be drawn with shapes and the pen tool in Keynote. Then lay them out according to the image's layout and process. A usable mechanism diagram is completed just like that.

After drawing the figure, I also go through the logic and order of the mechanism diagram myself to see whether anything needs improvement. In addition, for the core terms and processes, I check against the body text I have already verified to see whether there are differences, and only after ensuring consistency do I proceed to the next drawing task.

Some Drawing Tips

Drawing Layout

If your computer has an external monitor, drawing with a left-right split screen is a great experience; just keep this split-screen layout. If you don't have an external monitor, don't worry—drag the AI-generated image directly onto the PPT canvas and draw alongside it, and you'll still have a very good drawing experience.

Order of Drawing

When drawing, it is best to start from "local parts" rather than from the large section content.

Because drawing inevitably requires adjustments: if you draw framework-type and large layout content first, then when drawing small-scope content later, select-all and dragging will not be smooth enough (and layer issues will be involved), which instead brings a lot of inconvenience. Therefore, we can first draw the local content well at the beginning, trace each small part, and finally add the layout content—this is more efficient.

Layout and White Space

Reference: Make good use of these two "alignment" tools in Keynote, and slide layout can also be professional and textured

To avoid the mechanism diagram looking cluttered and crowded, we need to learn to control the white space around it. Before drawing, it is recommended to first use guides in the Keynote canvas to define uniform top, bottom, left, and right margins, and strictly limit all visual elements to the core area; the layout will immediately present a clean, spacious sense of breathing.

In this way, our PPT as a whole has white space, ensuring a certain degree of aesthetics.

Exporting Images

After completing the tracing work for each part and doing a bit of beautification, the image can be exported.

Please keep both the "editable source file" in Keynote and the final exported image file. The former is convenient for later modifications, and the latter is used for formal submission.

For scientific images, unless otherwise required, it is recommended to select all the figures drawn in the PPT software, "group" them, and directly paste and copy them into the corresponding position in the article. The clarity of such images is very high.

If the target journal's submission system requires uploading high-resolution images above 300 DPI separately (such as TIFF or PNG), exporting images directly from Keynote often cannot customize the resolution.

Here I share a native lossless conversion technique for macOS:

First, in Keynote choose "File -> Export To -> PDF" (select "Best" for image quality; vector information is preserved by default), then open the PDF with the system's built-in "Preview," click "File -> Export," and you can manually specify the resolution as 300 or 600 pixels/inch when exporting to TIFF/PNG.

After exporting, I also check whether the image is clear at the actual layout size, especially fonts, arrows, and smaller graphic elements. If there are problems, please return to the source file promptly to modify them.

Step 6: Pre-submission Verification

Writing the review is only the main part before submission, but there is still a distance to go before submitting.

We also need to modify the format and supplement information according to the journal's requirements. Organizing references is also a time-consuming task before submission. At that time, I found that some references imported into Zotero had incomplete DOI, volume, issue, or publication information. I would first use Google Scholar to find the corresponding paper, then click the "Cite" button for that reference to supplement the reference information in the review—volume number, authors, etc.

There are also details such as figure captions, table alignment, font size, etc., that require careful verification.

Traditional web-based AI requires repeatedly uploading documents and copying and pasting. But at the submission verification stage, I introduced a local AI Agent tool (using WorkBuddy as an example). It does not require dealing with technical issues such as network environments, IPs, or topping up accounts. It is a tool that works out of the box and is sufficient.

For example: after my paper has gone through a round of revisions, the order and format of the cited references may have some problems. So I set WorkBuddy's work project to my paper file and give it the instruction:

Title: How Graduate Students Can Use AI to Help Complete a Review Paper: My Research Writing Tools and Practical Experience

Body:

        I have finished a review paper. Please help me carry out a pre-submission format and reference cross-check.

Step 1: Read the copy of the paper in the current working directory 【insert file name.docx】. Do not modify the original file directly under any circumstances.

Step 2: Follow the Chinese national standard 【GB/T 7714-2015 numeric citation style】 (and the submission guidelines of the target journal 《【insert journal name】》) and check the following issues one by one:
1. Whether the in-text citation numbers and the reference list at the end correspond strictly one-to-one in order of appearance.
2. Whether there is any disconnection, such as "a citation appears in the text but the corresponding entry is missing at the end" or "an entry exists at the end but is never cited in the text."
3. Whether the author, title, journal name, year, volume/issue/page numbers, and DOI of the references have formatting omissions or obvious errors.
4. Whether the figure and table numbering throughout the paper (Figure 1, Table 1, etc.), the multi-level heading hierarchy, and the in-text references to them are consistent.
5. Whether there are common typesetting flaws such as mixed Chinese and English punctuation, or physical quantity units not italicized / not preceded by a space.

Step 3: First generate a "Verification Report" that outputs the following for each issue:
- Location (chapter/paragraph)
- Original text content
- Description of the problem
- Suggested revision
- Items that must be manually verified by the author

Step 4: Safety red line — do not insert any new literature citations on your own, do not tamper with the paper's scientific arguments, and do not rewrite the meaning of the main text in order to force a match with the references.

Step 5: Wait for me to review and confirm the revisions in the "Verification Report" before executing the corresponding replacements and rearrangement in the copy file.

Step 6: Keep the original file, and output the revised new file locally along with a complete changelog.
      

After executing the instructions, WorkBuddy automatically rearranges the in-text citations and the reference list at the end according to the numeric citation style, fills in missing DOI information, and saves the revised document in the local directory. Compared with traditional web-based large models, the biggest advantage of a local Agent is "targeted, lossless modification" — it changes only the citations and formatting errors I specify, and the layout style of the rest of the main text is completely unaffected, saving the tedious work of repeatedly uploading and copying.

After the review draft is complete, during the revision process for submission and resubmission, WorkBuddy can be a big help.

Final Thoughts

Looking back on this writing workflow, my biggest takeaway is this: what AI truly changes is not research writing itself, but the way we work when we do research writing.

In the past, we might have had to spend a lot of time on repetitive information organization, literature classification, and format adjustment. Now, with AI and some digital tools, we have the opportunity to make this work more efficient.

But tools are always just tools. Whether the topic is worthwhile, whether the research conclusions are reliable, whether there are contradictions among different pieces of literature, and how the final article should be expressed — these questions still require the author's own understanding and judgment. For those of us currently in graduate school, there may be no need to pursue a perfect research workflow from the very beginning. First find a question worth studying, then use suitable tools, and do it well step by step — that is already a very good start.

As a engineering graduate student myself, I often see peers on social media sharing the difficulties of graduate study — wanting to produce some results, yet suffering from the lack of guidance and a weak foundation, falling into anxiety, confusion, and self-doubt through day after day of trial and error. This state seems to have become the norm among graduate students, but that absolutely does not mean we can only accept it passively.

Has it always been so, and therefore it is right? — Lu Xun, "A Madman's Diary"

I hope this article can be of some help to students who are preparing a review or worrying about research writing.

• WorkBuddy official website

• Zotero Reports official documentation

• Make good use of these two "alignment" tools in Keynote, and your slide layout can also be professional and refined

Follow the Shaoshupai Xiaohongshu account and experience a wonderful digital life 🍃

Practical, easy-to-use genuine software, presented to you by Shaoshupai 🚀