Mini Literature Review with AI Tools

Use Elicit, Connected Papers, and NotebookLM to review research papers.

Introduction

30 Second Summary

Literature reviews are the foundation of any research project, but manually screening hundreds of papers, extracting data, and synthesizing themes can take weeks of painstaking work.

In this project, you will use a coordinated set of AI tools to conduct a complete mini literature review on scaffolded autonomy in engineering education. You will compress what typically takes weeks of manual effort into under an hour.

What You'll Build

A complete mini literature review covering 10-15 papers on how engineering educators balance learner autonomy with structured scaffolding during the high school to university transition, produced using a repeatable AI-assisted workflow.

By the end of this project, you'll have:

A visual citation graph showing the research landscape, foundational works, and emerging directions around your topic.
A structured extraction table with methodology, sample, scaffolding approach, and key findings for each paper.
A source-grounded thematic synthesis with identified research gaps, ready to expand into a publication.
Secret Mission: Generate an Audio Overview podcast of your findings in NotebookLM and critically evaluate what the AI gets right and wrong about your field.

Are there any prerequisites?

You should be familiar with academic literature review processes. You will also need an existing Elicit account, a Google account (for NotebookLM access), and Mendeley installed for reference management.

Connected Papers requires no account for the free tier, so you can use it directly in your browser.

Set Up Your AI Research Workspace

Before diving into your literature search, you need a coordinated workspace where each tool handles a distinct phase of the review process.

In this step, you will open and configure four tools: Elicit for semantic search and data extraction, Connected Papers for visual citation mapping, NotebookLM for source-grounded synthesis, and Mendeley for reference management.

In this step, get ready to:

Create project spaces in Elicit, NotebookLM, and Connected Papers.
Configure NotebookLM with a custom academic reviewer goal.
Set up a dedicated Mendeley collection for your review papers.

Create your project spaces

Each tool in your pipeline needs its own dedicated workspace so your findings stay organized across the entire review process.

Open Elicit in your browser at elicit.com and sign in to your existing account.
Click New session to create a fresh research session.
Leave this tab open. You will use it in Step 3 to run your first search.

You should see an empty Elicit session with a search bar ready for a research question.

Open a new browser tab and go to notebooklm.google.com.
Sign in with your Google account if prompted.
Click New notebook.
Title the notebook Scaffolded Autonomy Review.

You should see an empty notebook with a sources panel on the left and a chat panel on the right.

Open another browser tab and navigate to connectedpapers.com.
Confirm you can see the search bar on the homepage. No account is needed for the free tier.

Why Connected Papers over other mapping tools?

Connected Papers uses co-citation and bibliographic coupling to build its graphs from Semantic Scholar's index of 200M+ papers. This means it finds conceptually related work even when papers use different terminology.

The free tier gives you 5 graphs per month with no account required. That is more than enough for a single mini-review session.

Can't access one of the tools?

If Elicit asks you to create an account, use the same email you registered with previously. If NotebookLM is not available in your region, check that you are signed into a personal Google account (not a managed workspace account that may have restrictions).

Connected Papers should load without any sign-in. If the page does not load, try clearing your browser cache or using a different browser.

Help me troubleshoot tool access issues for my literature review workspace.

Configure NotebookLM for academic analysis

NotebookLM is strictly grounded to the sources you upload. It will only synthesize from those documents and cannot pull in outside information. This makes it ideal for literature reviews because it prevents hallucinated claims.

To get rigorous academic analysis rather than generic summaries, you need to set a custom goal that steers its responses.

In your Scaffolded Autonomy Review notebook, look at the chat panel on the right side of the screen.
Click the configuration icon (gear icon) at the top of the chat panel.
Find the Custom goal field.
Enter the following goal text:

Act as a rigorous education research reviewer who asks probing methodological questions, identifies gaps in evidence, and synthesizes findings with citations. When summarizing papers, always note the study design, sample characteristics, and limitations. When comparing across studies, highlight contradictions and areas of consensus.

What does this custom goal do?

The custom goal shapes every response NotebookLM generates in this notebook. Without it, NotebookLM provides general-purpose summaries. With this goal, it will automatically surface methodological concerns, cite specific papers, and point out where evidence is thin.

You can update this goal at any time if you want to shift the analysis focus later in your review.

Click Save to apply the custom goal.

Your notebook is now configured. When you upload papers in Step 5, NotebookLM will use this academic reviewer lens for all its responses.

Don't see the configuration icon?

The gear icon appears at the top of the chat panel, not the sources panel. Make sure you are looking at the right-hand side of the notebook interface. If you still do not see it, try refreshing the page.

Help me find the custom goal setting in NotebookLM.

Prepare Mendeley for your review collection

A dedicated collection keeps your review papers separate from the rest of your reference library. This makes it easy to track which papers are part of this specific review.

Open Mendeley on your computer.
In the left sidebar, click New Collection (or right-click in the collections area and select New Collection).
Name the collection Scaffolded Autonomy - AI Review.
Press Enter to confirm the name.

You should see the new collection appear in your left sidebar, empty and ready to receive papers.

Use this collection as your single source of truth

As you identify papers in Steps 3 and 4, add them to this collection immediately using DOIs or exports. This ensures your final reference list stays accurate and complete throughout the review process.

Your research workspace is ready. You have four tools configured, each handling a different phase of the review. Next up, you will define your research question and build a systematic search strategy to guide your literature discovery.

Define Your Research Question and Search Strategy

Your AI research workspace is ready. Now you need to give it direction.

A literature review is only as good as its search strategy. Vague questions return noisy results, and AI tools amplify that noise by surfacing hundreds of tangentially related papers. In this step, you will refine your broad interest into a precise, searchable research question and build the Boolean logic that will drive your searches across Elicit and traditional databases.

In this step, get ready to:

Formulate a focused research question with clearly identified constructs.
Build a Boolean search string with synonyms for each construct.
Define inclusion and exclusion criteria to scope your review.

Formulate a focused research question

A broad topic like "scaffolding in education" could match thousands of papers. You need to narrow it into a question that targets the specific intersection you care about.

The strongest research questions for literature reviews identify the intervention (what is being done), the population (who it is done to), and the context (where it happens). This structure makes your question directly translatable into search terms.

Open a new document (Google Docs, Word, or a plain text file) to record your search strategy.
At the top of the document, write your finalized research question:

"What instructional strategies scaffold learner autonomy during the transition from secondary to tertiary education in STEM/engineering contexts?"

Below the question, list your key constructs. These are the core concepts embedded in the question:

Scaffolding (the intervention).
Learner autonomy (the outcome of interest).
Transition pedagogy (the context, specifically secondary-to-tertiary).
First-year experience (the population stage).
Engineering education (the disciplinary context).

Why identify constructs separately?

Each construct becomes a column in your search matrix. By naming them explicitly, you can generate synonyms for each one independently. This is what transforms a single question into a comprehensive search that catches papers using different terminology for the same ideas.

Develop search terms and Boolean combinations

Researchers use different words for the same concepts. A paper about "fading support" is studying scaffolding, but it will not appear in a search that only uses the word "scaffolding." Your Boolean search string accounts for this by grouping synonyms with OR and connecting constructs with AND.

In your search strategy document, create a search matrix. For each construct, brainstorm 3-4 synonyms or related terms:

Scaffolding: scaffolding, "guided autonomy", "fading support", "structured freedom".
Autonomy: autonomy, "self-regulated learning", "self-directed learning".
Transition: transition, "first-year", "school-to-university", "secondary-tertiary".
Discipline: engineering, STEM, "higher education".

Combine these into a full Boolean search string and record it in your document:

(scaffolding OR "guided autonomy" OR "fading support" OR "structured freedom") AND (autonomy OR "self-regulated learning" OR "self-directed learning") AND (transition OR "first-year" OR "school-to-university" OR "secondary-tertiary") AND (engineering OR STEM OR "higher education")

How does this relate to Elicit?

Elicit uses semantic search, meaning it understands meaning rather than just matching keywords. You will enter your research question in natural language in the next step, not this Boolean string.

However, your Boolean string is still valuable. It documents your logic for methodological transparency. It also works in traditional databases (Scopus, Web of Science, ERIC) if you want to expand your search later.

Quote marks matter

Wrapping multi-word terms in quotation marks (like "self-regulated learning") forces databases to treat them as exact phrases. Without quotes, a database might match "self" and "learning" separately, returning irrelevant results.

Set inclusion and exclusion criteria

Even a well-crafted search will return papers outside your scope. Inclusion and exclusion criteria act as your screening filter. You define them now so that decisions during screening are consistent and defensible.

In your search strategy document, add a section titled "Inclusion Criteria" and list the following:

English-language peer-reviewed articles published between 2010 and 2026.
Empirical or design-based research (studies that collected and analyzed data).
Focused on instructional approaches (how teachers or institutions scaffold autonomy).

Add a section titled "Exclusion Criteria" and list:

K-12 only studies (must involve the transition to or experience of tertiary education).
Pure measurement or validation studies (papers that only develop a scale or instrument without reporting on an instructional intervention).

Why exclude measurement-only studies?

Your review focuses on what educators DO to scaffold autonomy. Papers that only validate a self-regulation questionnaire tell you how to measure the outcome, not how to produce it. These papers are useful for a different review question.

Save your search strategy document. You now have a complete, documented search protocol.

Keep your criteria visible

You will refer back to your inclusion and exclusion criteria in the next step when screening papers from Elicit and Connected Papers. Having them written down prevents scope creep as interesting but off-topic papers appear.

Your search strategy is locked in. You have a precise question, a documented Boolean string, and clear screening criteria. Next up, you will put this strategy to work by searching Elicit and mapping the research landscape with Connected Papers.

Discover and Map the Literature

You have a focused research question and a documented search strategy. Now it is time to find the actual papers that will form your evidence base.

In this step, you will use two complementary discovery approaches. Elicit finds papers through semantic search (meaning over keywords), while Connected Papers reveals the citation neighborhood around your most relevant result.

In this step, get ready to:

Run a semantic search in Elicit and identify your most relevant papers.
Build a citation graph in Connected Papers and discover foundational and emerging work.
Consolidate a list of 10-15 papers and add them to your Mendeley collection.

Search with Elicit

Elicit searches across 125M+ academic papers using natural language. Instead of relying on exact keyword matches, it understands the meaning behind your question and surfaces conceptually relevant work.

In your open Elicit research session, click the search bar at the top of the page.
Enter your research question in natural language:

What instructional strategies scaffold learner autonomy during the transition from secondary to tertiary STEM education?

Press Enter to run the search.
Review the results page. Elicit displays paper titles, authors, publication years, and AI-generated summaries of each paper's relevance to your question.
Scan the first page of results and identify 3-5 papers that are highly relevant to your research question.
Note the titles of these papers somewhere accessible (a text file, sticky note, or your search strategy document).

What makes a paper 'highly relevant'?

Look for papers that directly address instructional strategies (not just student experience surveys) in a STEM or engineering context during the school-to-university transition.

Prioritize empirical or design-based research over purely theoretical pieces. Your inclusion criteria from Step 2 guide this decision.

Pick your strongest result carefully

From your 3-5 relevant papers, decide which single paper is the most central to your research question. You will use this paper as the seed for Connected Papers in the next substep. A well-chosen seed produces a more useful citation graph.

Map the landscape with Connected Papers

Connected Papers builds visual citation graphs using co-citation and bibliographic coupling. This means it finds papers that are frequently cited together or cite the same sources, revealing conceptual neighbors that keyword search alone would miss.

Switch to your Connected Papers browser tab at connectedpapers.com.
In the search bar, enter the title of your most relevant paper from the Elicit results.
Select the correct paper from the dropdown suggestions.
Click Build a graph.

Connected Papers generates a visual network of related papers. This takes a few seconds to render.

Examine the graph that appears. Larger nodes represent more-cited papers (foundational works). Darker nodes represent more recent publications.
Look for clusters of tightly connected nodes. These clusters represent sub-topics or methodological approaches within your research area.
Click the Prior Works tab in the left panel. This shows foundational papers that the field is built upon.
Click the Derivative Works tab. This shows more recent papers that built on or extended your seed paper.
Identify 5-8 additional papers that appear central to the graph. Focus on large nodes (highly cited), papers that bridge clusters, and recent derivative works.
Note the titles of these papers alongside your Elicit results.

Why does Connected Papers find different papers than Elicit?

Elicit uses semantic similarity to your question. Connected Papers uses citation relationships from Semantic Scholar's index of 200M+ papers. A paper might never mention your exact keywords but still be central to the conversation because it is frequently co-cited with your seed paper.

This is why combining both approaches gives you better coverage than either tool alone.

Graph not showing your paper?

Connected Papers accepts a title, DOI, URL, or ArXiv ID. If the title search does not find your paper, try pasting the DOI instead.

Very recent papers (published in the last few months) may not yet appear in Connected Papers' index. If this happens, try your second-most relevant paper from Elicit as the seed instead.

Remember that the free tier allows 5 graphs per month. Choose your seed paper carefully to avoid using up graphs on less relevant results. Help me find my paper in Connected Papers.

Consolidate your paper set

You now have papers from two complementary sources. Combining them gives you a set that includes both semantically relevant work (from Elicit) and structurally connected work (from Connected Papers).

Combine your Elicit results (3-5 papers) and Connected Papers discoveries (5-8 papers) into a single list.
Remove any duplicates that appeared in both sources.
Verify your total is in the range of 10-15 papers. If you have fewer than 10, revisit Connected Papers and check the Derivative Works tab for additional recent studies.
Open Mendeley and navigate to your Scaffolded Autonomy - AI Review collection.
Add each paper to the collection. Use DOIs where available for accurate metadata import. You can also export papers directly from Elicit or Connected Papers if they offer a citation export option.
Verify that all 10-15 papers appear in your Mendeley collection with correct titles and authors.

Why both tools instead of just one?

Elicit excels at finding papers that answer your specific question, even if they use different terminology. Connected Papers excels at revealing the broader conversation around a topic, including foundational works you might not find through search alone.

Together, they ensure you capture both keyword-relevant and conceptually-related work. This combined approach is what makes an AI-assisted review more comprehensive than manual database searching.

Flag papers for full-text reading

As you add papers to Mendeley, make a mental note of which ones you have full PDF access to through your institution. You will need these PDFs in Step 5 when you upload sources to NotebookLM for synthesis.

Your paper collection is assembled and organized. Next up, you will use Elicit's extraction tools to pull structured data from each paper into a table you can analyze.

Screen and Extract Structured Data with Elicit

You have a curated list of 10-15 papers sitting in your Mendeley collection. Now you need to transform those papers from a flat list into a structured evidence base you can analyze for patterns.

In this step, you will use Elicit's AI-powered extraction to pull structured data points from each paper automatically. Raw papers become an organized table of evidence ready for thematic synthesis.

In this step, get ready to:

Build a structured extraction table in Elicit with custom columns for your key data points.
Verify AI-extracted data against original papers for accuracy.
Export your evidence base and flag papers needing deeper reading.

Build an extraction table in Elicit

Defining specific data points upfront forces you to compare papers on the same dimensions. This prevents the common pitfall of summarizing each paper individually without ever finding cross-cutting patterns.

In your Elicit session, add your remaining papers from the consolidated list by clicking Add papers and entering each paper's title or DOI.
Continue until all 10-15 papers from your Mendeley collection appear in the Elicit table.

Use DOIs for accuracy

When adding papers manually, paste the DOI instead of typing the title. DOIs are unique identifiers that guarantee Elicit finds the exact paper you want.

Click the Column button at the top of the table to add a custom column.
Name your first column Study Design/Methodology and add instructions describing what you want extracted (e.g., "What research methodology did this study use? Include whether it is qualitative, quantitative, or mixed methods, and name the specific design such as RCT, case study, or design-based research.").
Click Save to create the column.
Repeat the process for your second column: Sample Size and Population.

Free tier column limit

Elicit's Basic (free) plan allows 2 custom columns per table. Since you need 6 columns total, you will work in batches of 2 columns at a time.

After extracting and recording data from your first 2 columns, delete them and create the next pair. Keep a running record of all extracted data in a separate spreadsheet or document as you go.

Once your first 2 columns are populated, copy the extracted data into a separate spreadsheet (Google Sheets or Excel).
Delete those columns, then create your next pair: Scaffolding Approach Used and Autonomy Outcome Measures.
Copy that data to your spreadsheet, then repeat for the final pair: Key Findings and Limitations.

What do these columns capture?

Study Design/Methodology: The research approach (RCT, quasi-experimental, case study, design-based research).
Sample Size and Population: How many participants, what year level, and which discipline.
Scaffolding Approach Used: The specific instructional strategy (fading prompts, guided choice, metacognitive supports, peer scaffolding).
Autonomy Outcome Measures: How the study measured whether students became more self-directed.
Key Findings: The main results and effect sizes reported.
Limitations: What the authors identified as weaknesses or constraints of their study.

Run the extraction and verify accuracy

Elicit's AI populates your columns automatically by reading paper abstracts and full texts (where available). This saves enormous time, but AI extraction is not perfect. Your expert judgment is the final quality gate.

After creating each pair of columns, wait for Elicit to populate the cells. This typically takes 10-30 seconds per paper.
Click on any cell in the extraction table to view the supporting quotes from the source paper. Elicit shows the exact passage it used to generate the answer.
Select 3-4 papers you are most familiar with and compare the extracted data against your own reading of those papers.
Check specifically for: missing nuance (e.g., a mixed-methods study labeled as purely quantitative), conflated constructs (e.g., self-efficacy reported as autonomy), and blank cells where data should exist.

What accuracy should you expect?

Independent testing shows Elicit achieves 94-99% accuracy on well-structured empirical papers. Accuracy drops for theoretical or qualitative papers where key information is distributed across many paragraphs.

If Elicit only has access to the abstract (not the full PDF), extraction will be limited to what the abstract contains. You can check this by looking for a PDF icon next to each paper in your table.

For any cell where the extraction seems wrong or incomplete, click the cell and manually edit it based on your reading of the paper.
Note which papers had incomplete extraction. These are the ones you will flag for full-text reading later.

Seeing blank or vague cells?

Blank cells usually mean Elicit could not find the relevant information in the abstract. This happens when papers do not report key details in their abstract (common for sample size and limitations).

Try uploading the PDF directly to Elicit if you have library access. Elicit can then search the full text for your answer. If a cell says something generic like "not reported," flag that paper for manual full-text reading.

Help me troubleshoot blank cells in my Elicit extraction

Export and organize your extraction data

Your extraction table is your primary evidence artifact for this review. You need it in a portable format you can reference during synthesis and include in your final document.

If Elicit offers an Export button above your table, click it and select CSV as the format. Save the file to your Desktop.
If export is not available on your plan, open a new Google Sheet or Excel spreadsheet. Create column headers matching your 6 extraction categories. Copy the data from each Elicit column into the corresponding spreadsheet column. Save the file as a .csv file.
Review your completed spreadsheet. For each paper, check whether the extraction captured enough detail to support synthesis.
Flag papers that need full-text reading by adding a new column called Needs Full-Text Review and marking Yes for any paper where Elicit could not extract sufficient detail from the abstract alone.

What makes a paper need full-text reading?

Flag a paper when: the scaffolding approach column is vague or generic, the sample details are missing, or the key findings cell contains only the conclusion sentence from the abstract without effect sizes or specific results.

These papers still belong in your review. They just need manual reading to complete their data row.

Save your column definitions

Click the arrow next to any column name and select Save to column library before deleting it. This saves your column definitions as presets you can reuse in future Elicit sessions without retyping the instructions.

Your structured extraction table is ready. In the next step, you will upload your source papers to NotebookLM and use its source-grounded AI to synthesize themes across your entire evidence base.

Synthesize Themes Using NotebookLM

Your extraction table has structured data from 10-15 papers, but individual data points don't tell a story on their own. The real value of a literature review comes from synthesis: identifying patterns, tensions, and gaps that span multiple studies.

NotebookLM's source-grounded analysis is ideal for this phase. It synthesizes only from your uploaded documents and cites specific sources for every claim, preventing the hallucination problem that makes general-purpose LLMs risky for academic work.

In this step, get ready to:

Upload your paper PDFs as sources in your NotebookLM notebook.
Run targeted synthesis prompts to identify themes, tensions, and gaps.
Generate a structured thematic summary with cited sources.

Upload your sources to NotebookLM

NotebookLM can only synthesize from documents you explicitly provide. You need to upload the actual PDFs so the tool has full-text access, not just abstracts.

Download PDFs of your 10-15 papers from your university library.
Use your Mendeley collection Scaffolded Autonomy - AI Review as your checklist to confirm you have a PDF for each paper.
In your NotebookLM notebook Scaffolded Autonomy Review, click the + button in the Sources panel on the left.
Select PDF as the source type.
Upload all your paper PDFs. You can select multiple files at once.
Wait for NotebookLM to process each document. A checkmark or status indicator appears next to each source once processing is complete.

Why upload PDFs instead of just abstracts?

NotebookLM is strictly grounded to uploaded sources. It can only cite and synthesize information from the full text you provide.

With full PDFs, the tool can reference methodology sections, discussion points, and nuanced findings that never appear in an abstract. This makes its synthesis far richer than abstract-only analysis.

NotebookLM source limits

The free tier supports up to 50 sources per notebook. Your 10-15 papers are well within this limit. NotebookLM also accepts Google Docs, web URLs, and pasted text if any PDF is unavailable.

Once all sources show as processed, you should see each paper listed in the Sources panel with its title visible.

Sources stuck on processing?

If a PDF stays in a processing state for more than a minute, check that the file is under 200MB and is a text-based PDF (not a scanned image). Try re-uploading the file.

If the issue persists, paste the paper's web URL as an alternative source type instead of the PDF. Why won't my PDF upload to NotebookLM?

Conduct source-grounded thematic analysis

With all your papers loaded, NotebookLM can now answer questions by drawing on the full text of every source simultaneously. Your custom academic reviewer goal steers it toward rigorous, citation-backed responses.

In the chat panel of your NotebookLM notebook, type this first synthesis prompt:

Across these sources, what are the main approaches to scaffolding student autonomy in the first year? Cite specific studies for each approach you identify.

What does this prompt target?

This first prompt maps the landscape of approaches in your literature set. By requesting specific citations, you ensure every claim traces back to a source you can verify.

Review the response. NotebookLM should list several distinct approaches with inline source references (usually author name or document title).
Once you have the first response, type this second prompt to identify tensions:

What tensions or contradictions exist between these approaches? Where do the authors disagree about how much structure versus freedom is appropriate?

Review the response for opposing viewpoints or methodological disagreements between papers.
Type this third prompt to surface gaps in the evidence base:

What gaps in the evidence base do these papers collectively reveal? What questions remain unanswered, and what populations or contexts are underrepresented?

Why three separate prompts instead of one?

Each prompt targets a different analytical layer: descriptive mapping, critical analysis, and gap identification. Separating them produces more focused, detailed responses than a single broad request.

This mirrors how experienced reviewers work through literature: first understanding what exists, then interrogating tensions, then identifying what is missing.

You should now have three substantive responses in your chat history, each citing specific papers from your uploaded sources.

Responses seem thin or generic?

If NotebookLM gives short or vague answers, check that all your sources finished processing (unprocessed PDFs are invisible to the chat). Also verify your custom academic reviewer goal is still active by clicking the configuration icon in the chat panel.

Try rephrasing your prompt with more specific terminology from your research question, such as naming specific scaffolding types or citing a specific author. How can I get more detailed responses from NotebookLM?

Generate a structured thematic synthesis

Your three exploratory prompts revealed the key patterns in your literature. Now you will ask NotebookLM to organize everything into a coherent thematic structure you can use in your review document.

Type this synthesis prompt into NotebookLM:

Produce a structured thematic summary of the scaffolding approaches across these sources. Organize by themes: (1) fading scaffolds, (2) guided choice architectures, (3) metacognitive supports, and (4) peer-based scaffolding. For each theme, cite the specific sources that provide evidence. End with a section on research gaps and future directions.

What does this prompt do?

This prompt asks NotebookLM to reorganize everything it found into a publication-ready thematic structure. The four theme categories come from your earlier exploratory prompts.

By naming the themes explicitly, you guide the AI toward the structure your review needs rather than letting it impose its own arbitrary organization.

Review the structured output. Each theme should have cited evidence from multiple papers.
Select all the text in NotebookLM's response and copy it to a new document (Google Doc, Word, or plain text file).
Title this document Draft Thematic Synthesis.
Read through the copied synthesis and highlight any claims that feel overstated or that you cannot immediately recall from your reading. Mark these with a [VERIFY] tag.

Why flag claims for verification?

NotebookLM is grounded to your sources, which means it cannot hallucinate facts from outside your papers. However, it can still misinterpret a finding, overstate a conclusion, or attribute a claim to the wrong source.

Your expert knowledge of engineering education is the quality filter. Flagging uncertain claims now saves you from publishing errors later.

Refine themes that feel incomplete

If one theme has fewer citations than others, ask a follow-up prompt like: "Which sources discuss [theme name] specifically? Are there related findings I should include?" NotebookLM's 50 chats per day limit gives you plenty of room to iterate.

Your draft synthesis document should now contain a structured thematic summary with four main themes, cited sources under each, and flagged claims ready for verification.

NotebookLM missed a theme or merged themes together?

If the output collapses two themes into one (e.g., treating metacognitive supports and fading scaffolds as the same thing), ask a clarifying follow-up: "Separate your discussion of metacognitive supports from fading scaffolds. How do these differ in approach and evidence?"

If a theme returns no evidence, it may not be represented in your particular paper set. Note this as a potential gap rather than forcing the theme. Help me refine my NotebookLM synthesis themes

You now have a source-grounded thematic synthesis that would have taken days to produce manually. Next up, you will assemble this synthesis alongside your extraction table and methodology notes into a complete mini-review document.

Produce Your Mini-Review Document

Your thematic synthesis from NotebookLM gives you the raw material. Now you need to shape it into a structured academic document that a reader (or reviewer) can follow.

AI-generated synthesis is a starting point, not a final product. In this step, you will draft the mini-review structure, populate it with your artifacts, and critically verify every claim against your original papers.

In this step, get ready to:

Draft a mini-review document with Introduction, Methods, Thematic Findings, Research Gaps, and Implications for Practice sections.
Integrate your Elicit extraction table, Connected Papers insights, and NotebookLM synthesis into the document.
Verify every claim against original sources and flag any AI confabulations.

Draft the review structure

A mini-review follows a predictable academic structure. Setting up the skeleton first means you can slot your existing artifacts into the right places without rewriting from scratch.

Open a new document in your preferred writing tool (Google Docs, Word, or any academic writing software).
Create the following section headings in this order:

Introduction (research question and scope).
Methods (your AI-assisted search and selection process).
Thematic Findings (organized by the themes NotebookLM identified).
Research Gaps (what the collective evidence does not yet address).
Implications for Practice (what educators should do differently based on the findings).

Under the Introduction heading, paste your finalized research question: "What instructional strategies scaffold learner autonomy during the transition from secondary to tertiary education in STEM/engineering contexts?"
Add 1-2 sentences explaining why this question matters for engineering education practice.
State your scope: the number of papers reviewed, the date range (2010-2026), and the focus on empirical or design-based research.

Why this structure?

This five-section structure follows the conventions of a scoping review or mini-review in education research. Reviewers and readers expect this flow because it mirrors the logic of your investigation: what you asked, how you found papers, what you found, what is missing, and what it means.

Integrate your artifacts

Each section of your mini-review draws on a different artifact you have already produced. The Methods section describes your process. The Findings section expands on NotebookLM's synthesis with your expert interpretation.

In the Methods section, describe your three-tool workflow in this order:

Elicit for semantic search and structured data extraction (mention the natural language query you used and the custom columns you created).
Connected Papers for visual citation mapping and discovery of conceptually-related work (mention your seed paper and how you identified additional papers from the graph).
NotebookLM for source-grounded thematic synthesis (mention the custom academic reviewer goal and the targeted prompts you used).

Also describe your inclusion/exclusion criteria in the Methods section: English-language peer-reviewed articles, 2010-2026, empirical or design-based research focused on instructional approaches.
Insert your Elicit extraction table CSV as an appendix or as a formatted summary table at the end of the document.

Summary table vs. full appendix

If your mini-review is short (under 3,000 words), a summary table with Author, Year, Method, Sample, and Key Finding columns works well inline. For longer documents, put the full extraction table in an appendix and reference it from the Findings section.

In the Thematic Findings section, create subheadings for each theme NotebookLM identified. Your themes should include: fading scaffolds, guided choice architectures, metacognitive supports, and peer-based scaffolding.
For each theme, copy the relevant synthesis from your NotebookLM output as a starting point.
Expand each theme with your own expert interpretation. Add critical analysis: Which studies have stronger methodologies? Where do findings conflict? What contextual factors might explain differences?
In the Research Gaps section, draw from NotebookLM's response to your "What gaps in the evidence base do these papers collectively reveal?" prompt. Add your own observations about missing populations, understudied contexts, or methodological limitations you noticed.
In the Implications for Practice section, write 2-3 actionable recommendations for engineering educators based on your findings. Each recommendation should cite specific papers from your review.

What does expert interpretation look like?

NotebookLM reports what the papers say. Your job is to evaluate how strong the evidence is, whether the studies actually support the claims being made, and how the findings connect to broader theory (like Vygotsky's zone of proximal development or Kift's transition pedagogy framework).

Ask yourself: Would I cite this study to support a grant application? If not, note its limitations explicitly in your Findings section.

Critical review and quality check

AI tools accelerate literature review, but they can also introduce errors. NotebookLM is grounded to your uploaded sources, which reduces hallucination risk. However, it can still misinterpret findings or overstate conclusions.

Read through your entire document from start to finish.
For every factual claim in the Thematic Findings section, cross-reference it against your Elicit extraction table. Check that the study design, sample size, and findings match what you are reporting.
Flag any assertions where the AI may have overgeneralized or conflated two studies. Highlight these in a different color or add a comment for later revision.
Check that every paper you cite in the document appears in your Mendeley collection "Scaffolded Autonomy - AI Review." Remove any citations from the document that do not have a matching Mendeley entry.
Verify that your Mendeley collection does not contain papers you never ended up citing. Clean up any papers that fell outside your inclusion criteria.

What counts as AI confabulation in this context?

NotebookLM is strictly grounded to your uploaded sources, so it will not invent papers that do not exist. However, it can misattribute a finding to the wrong paper, overstate a conclusion that was tentative in the original, or merge findings from two different studies into one claim.

Your extraction table is the verification tool here. If a claim in your synthesis does not match the Key Findings column for that paper, flag it.

Correct any flagged errors by returning to the original PDF (either in NotebookLM or your local files) and rewriting the claim to accurately reflect the source.
Save your completed mini-review document.

Your document should now have all five sections populated, your extraction table attached as an appendix, and every claim verified against the original papers. Your expert knowledge of engineering education is the final quality filter that no AI tool can replace.

Finding claims that don't match your sources?

If you find multiple misattributed claims, return to NotebookLM and re-ask your synthesis question with more specific instructions. For example: "For each claim, name the single paper it comes from and quote the relevant sentence." This forces more precise citations you can verify quickly.

help me verify whether NotebookLM has correctly attributed findings to the right papers in my literature synthesis

Secret mission

Generate and Critically Evaluate an Audio Overview

You've synthesized findings by reading and prompting. But what happens when an AI tries to discuss your research out loud? Use NotebookLM's Audio Overview to generate a podcast-style conversation about your papers, then critically evaluate what it gets right and wrong about your field.

Clean Up Your Resources

This project uses only free-tier tools and browser-based services, so there are no ongoing costs. Decide whether to keep your resources available for continued research, pause by closing sessions, or delete them entirely.

Resources you used:

Elicit research session (with extraction table and saved papers).
NotebookLM notebook ("Scaffolded Autonomy Review" with uploaded PDFs and chat history).
Connected Papers graphs (up to 5 free graphs per month).
Mendeley collection ("Scaffolded Autonomy - AI Review").
Downloaded PDFs stored locally on your computer.
Exported CSV extraction table.
Draft mini-review document.

Keep everything running

No action needed. Choose this if you plan to continue expanding your literature review or want to revisit your synthesis later.

Your Elicit session, NotebookLM notebook, and Mendeley collection will remain accessible whenever you log back in.
Your Connected Papers graphs remain viewable even after your monthly limit resets.
Keep your downloaded PDFs and CSV export in a dedicated folder for easy access.

Pause - I'll come back to this later

Close your active browser tabs and sessions to free up memory. Your data stays saved in each platform.

Close your Elicit browser tab. Your session and extraction data are saved automatically.
Close your NotebookLM tab. Your notebook, uploaded sources, and chat history persist in your Google account.
Close the Connected Papers tab. Previously built graphs remain accessible when you return.
Move your downloaded PDFs, CSV export, and draft document into a clearly named folder on your Desktop.

Delete - I don't want to use this again

Remove all project resources and start fresh. This cannot be undone for most platforms.

In Elicit, open your research session and delete it from your session list.
In NotebookLM, open the "Scaffolded Autonomy Review" notebook, click the three-dot menu, and select delete.
In Mendeley, right-click the "Scaffolded Autonomy - AI Review" collection and select delete.
Delete the downloaded PDFs, CSV export, and draft mini-review document from your computer.
Connected Papers graphs expire naturally and do not need manual deletion.

That's a Wrap!

Nice work! You just compressed what typically takes 2-3 weeks of manual literature review into under an hour using a coordinated AI tool pipeline.

You now have a publication-ready mini-review section, a curated reference collection, and a repeatable workflow you can apply to any future research question.

Here's what you learned:

How to combine Elicit for semantic discovery and structured extraction, Connected Papers for visual citation mapping, and NotebookLM for source-grounded synthesis into a repeatable literature review pipeline.
How to verify and quality-check AI-extracted data against original sources to catch inaccuracies and confabulations.
How to use NotebookLM's custom goals to steer AI synthesis toward rigorous academic standards and produce structured thematic analysis with inline citations.
Secret Mission: Critically evaluated an AI-generated Audio Overview of your research findings to identify where AI adds value and where it flattens disciplinary nuance.

Ready to quiz yourself?