Claude Code for researchers

You read more than you write down, and you write down more than you can ever find again. Here is what changes when the folder of papers becomes a system that remembers where an idea came from.

Research has a specific version of the filing problem. It is not that you lack notes. It is that the notes and the source are two different objects living in two different places, and the link between them decays the moment you close the PDF. Six weeks later you remember the claim, not the citation, and you spend an hour re-finding a paper you have already read once.

The instinct is to build a better reading system: a reference manager, a highlighting workflow, a folder structure. Those solve storage. They do not solve the actual failure, which is that nobody, and no app, is doing the work of connecting today's paper to last month's note while you are still holding the thread in your head. That is the same upkeep problem covered from a workflow angle on Zettelkasten with AI: automatic linking, and it is worth reading if atomic notes and emergent connections are already how you think.

The cost compounds over the course of a project rather than showing up all at once. Early on, re-finding a source costs a few minutes and feels like a minor annoyance. By the time a dissertation or a multi-year research program is underway, the same habit has cost days spread across the whole project, all of it hidden inside what looks like normal reading time. Nobody budgets for it because it never shows up as a single expensive event, only as research that consistently takes longer than it should.

What actually changes

The useful version of this is not a smarter search bar. It is an assistant that can read the paper alongside you, pull out the claim, and file it against the question or the project it actually belongs to, the same day, instead of a someday-pile that never gets sorted. Reading capture without the copy-paste is the whole point of save articles for AI: a reading capture workflow, built for exactly this habit of reading more than you file.

The other half is the write side. A literature review, a lit-scan for a proposal, a background section for a paper, all of it is easier when the notes you already filed can be pulled back out in the shape you need, with the source still attached. That only works if the filing happened in a form you can actually read later, which is the argument this whole site keeps making: plain files you own, not a database you have to trust.

Where this differs from a search tool over your notes

Tools that summarize a document you already have open, or answer a question from notes you already wrote, are useful and worth knowing the limits of. NotebookLM vs Obsidian for research notes covers one specific version of that trade-off directly. The gap they all share is the same one: none of them decide, on their own, that today's reading belongs with last year's project. That is a filing decision, and it is the one worth automating first.

If you already work close to code or a technical stack, the adjacent case for that side of research work is on the consultants page, which covers the same juggling problem from a client-facing angle rather than a source-facing one.

Working with collaborators and advisors

Research is rarely solitary for long. A co-author, an advisor or a lab mate asks where a claim came from, and the honest answer is often "I remember reading it somewhere." Filed correctly, that question stops being a memory test. The source, the claim and the reasoning are already attached to the project, which means handing a collaborator the current state of the thinking takes minutes instead of a reconstruction session before every meeting. The same applies to grant proposals and progress reports, where the actual bottleneck is rarely the writing, it is finding the evidence you already gathered months ago and forgot you had.

Getting started

Start with whatever you are reading this week. Point the system at it, let it file the claim against the right project, and see whether you can find it again in a month without remembering the file name. That single test tells you more than any feature list.