How to run a literature review with Claude or ChatGPT, from your own notes

Ask Claude or ChatGPT for "a literature review on X" and you get something that reads like a literature review. Confident topic sentences. A tidy arc from foundational work to open questions. Citations in the right shape.

Then you check the citations. Half are papers you never read. Some do not exist. The ones that are real are summarized from the abstract, not from the thing the paper actually argues. And not one of them is the messy, important paper you spent last Tuesday annotating.

The problem is not that the model cannot synthesize. It synthesizes fine. The problem is what it is synthesizing from: a blurry average of its training data, plus whatever you managed to paste into this one thread before the context window filled up. It has never seen your reading. So it does a literature review of the literature in general, when what you needed was a literature review of yours.

The fix is to change the source, not the prompt

You can spend an afternoon engineering the perfect prompt and it will not fix this. A better prompt over the wrong source material gives you a better-worded version of the same problem. This is the retrieval problem, not a generation problem: the model is good enough, it just cannot see your work.

So give it your work. Keep your literature notes, one note per paper with your summary and your commentary, in a knowledge base your assistant can read directly. Then the synthesis is grounded in the papers you actually read, and every claim traces back to a note you can open and check.

That is the whole move. The rest is setup.

Step 1: Put one note per paper somewhere your AI can read it

You almost certainly already make these notes. The key finding. The method you might borrow. The line you disagreed with. The connection to that other paper. Today they live in the margins of a PDF, a Zotero annotation, or your head.

Move the thinking into notes. Not the PDF: your reference manager can keep that. The note holds what you would say about the paper if a colleague asked: what it found, whether you believe it, and how it connects to everything else you have read.

A note per paper is enough to start. A good one looks like:

# Argote & Ingram (2000): Knowledge transfer in organizations

**Finding:** Knowledge transfer is harder across organizational
boundaries than within them; embeddedness in the source context
is both what makes knowledge valuable and what makes it stick.

**My take:** Useful frame for the distributed-teams question, but
the "boundary" is treated as binary. Real teams have gradients.

**Connects to:** Szulanski (1996) on stickiness; tension with
Hansen (1999) on weak ties moving codified knowledge faster.

The "My take" and "Connects to" lines are what no reference manager and no training set has. They are the literature review, in pieces, before it is written.

Step 2: Connect Claude or ChatGPT over MCP

Hjarni is a knowledge base with a built-in MCP server, which is the protocol both Claude and ChatGPT use to read external tools. Connect it once and your assistant can search and read every note, in every conversation, without you pasting anything.

Setup takes a few minutes: connect Claude or connect ChatGPT, and the step-by-step wiki guide walks through it end to end. Both clients read the same notes, so you can draft in Claude and fact-check in ChatGPT against the identical source.

The difference from pasting: this is persistent and it scales. Thirty papers do not have to fit in one context window, because the model fetches the handful it needs for the question in front of it. And the work does not vanish when the thread ends. Add a paper next week and the next conversation already includes it.

Step 3: Ask for synthesis, and demand its sources

Now the prompts that used to produce confident fiction produce grounded work:

• "What are the main tensions in my notes on knowledge transfer barriers?"
• "Compare the three methods papers in my Literature Notes folder. Where do they disagree on measurement?"
• "Draft a literature review section from my annotations on distributed teams. Cite which note each point comes from."

That last instruction is the one that matters. Tell the assistant to attribute every claim to the note it came from. When the source is your own annotated note instead of a half-remembered abstract, you can click through and check it. Hallucinated citations are a symptom of guessing from training data; when the model is reading your notes, there is nothing to guess.

Make it a standing rule rather than retyping it every time. An AI instruction like "Always cite which of my notes you are drawing from, and say so explicitly when something is not in my notes" turns "trust me" into "check me" for every conversation.

Why this beats a smarter model

The instinct when an AI literature review disappoints is to wait for a smarter model. But the next model will still not have read your papers. It will be a more eloquent stranger to your work.

A reference manager owns your citations and PDFs, and it should keep doing that. What it does not own is your thinking: the annotations, the connections, the open questions. Give that thinking a home your assistant can read and you change what the model is for: not a search engine that returns plausible-sounding summaries, but a reader that has been through your stack and can help you think across it.

An AI that knows what you've read can help you think. One that doesn't is just a search engine.

Start with a structure

You do not have to design the folders yourself. The Knowledge Wiki template gives you a research wiki with sources, topics, open questions, and a changelog, plus the AI instructions that keep it tidy. Paste the template link into Claude or ChatGPT and it builds the initial structure for you; then it is just one note per paper from there.

The first review is the slow one, because you are writing the notes as you read. Every review after that starts from a corpus that already knows what you think. That is the part no model ships with, and the part that makes the synthesis yours.

If you want the longer version of the argument for researchers specifically, see Hjarni for researchers.