A system for literature reviews

A literature review system is a named queue of papers, a three-pass filter so most PDFs die early, and an export of claims you can write from without reopening every file. A folder named "to read" is not a process. It is a graveyard with better typography.

How do you run a literature review without drowning?

Cap the list, name it as a question, filter hard, RSVP what survives, and export marks before you write. You need three things: a queue with an order, a way to get through each item without rereading it three times, and a place the insights go so they are not trapped in the last scroll position.

How should you name a literature review list?

Name the review as a question, not a topic. "Does retrieval augmentation actually reduce hallucination on long-context benchmarks?" is a list. "AI papers" is a junk drawer.

Cap it. A first pass of 8–12 papers is a week of work if you actually finish them. Add more after you know what is missing. Infinite lists never start.

What order should papers go in?

Put the survey or canonical paper first, then the papers that argue with it, then the ones you are only reading so you can say you did. That order is a Readlist — a playlist for documents. You read them in sequence. When one is done, you are prompted into the next instead of returning to the folder to negotiate with yourself.

How do you filter papers before a full read?

Use a three-pass method. Most items should die in pass 1 (title, abstract, conclusion). The ones that survive get a real RSVP pass with marks on claims, methods, and limitations.

This is where speed is honest: you are not racing a paper you will cite in a footnote. You are refusing to spend forty minutes on a paper the abstract already ruled out.

What should you capture from each paper?

Capture claims, methods, and limitations as marks that still make sense with the PDF closed. After review you want a list like:

• "Claim: retrieval helps on HotpotQA, not on NarrativeQA — p.12"
• "They fine-tune the retriever; the generator is frozen — p.4"
• "No human eval, only automatic overlap — limitation"

Export as Markdown into the notes system you already have. The reading tool should not become a second brain you never open. It should fill the one you use.

When is a literature review done?

A review is done when you can write the related-work paragraph without reopening every PDF. If a hole remains, add two papers, not twenty. Then run the same loop.

Track the boring numbers: papers in the list, papers finished, marks kept. That is the same idea as a training log. You would not run a lifting program without knowing which day it is. Reading is a practice too.

What should you do this week?

This week: one question, at most ten PDFs, one sitting of pass 1, then RSVP and export.

1. Create a Readlist named after the question.
2. Drop in no more than ten PDFs.
3. Pass 1 on all of them in one sitting. Remove the duds.
4. RSVP and mark the rest, one after another.
5. Export marks. Write the paragraph while it is still warm.