My materials were scattered a year ago: course slides in the Downloads folder, translations on the desktop, notes in the notes app, and formatting requirements buried in group chats. When I needed to write something, I remembered"I've seen it", but I couldn't find it.
Now they are in one library. This post is about how four types of materials get in, how I ask questions after they are in, and how I get back to the source.
Four types of materials, three ways of handling them
Course slides and lecture notes. This type is the easiest: I put the whole thing in. They are my"background layer", and their use is to save me from looking things up in textbooks again when I ask a basic question.
Literature I have read (including translated drafts). I only put in what is related to my topic, not the complete set. My criterion is"will I cite this paper"——if I will cite it, it goes in; if I have only read it, I record it in a literature log and it does not go into the library.
My own notes and experiment records. This is the most valuable category in my library, and the only one that is not allowed to be lost. I name them by"date + topic", not by using"inspiration""quick notes"as a naming convention——the latter becomes garbage after three months.
Specification documents (templates, formatting requirements, procedures). I do not look at this type often, but every time I do, it is urgent. After putting them in, I no longer ask"what does the school require for figure captions", but rather"what are the figure caption requirements in my materials".
My three collection rules
- Unified naming rules, otherwise I cannot find them when asking questions.
- Granularity control: whole files go in; I do not put excerpts into the library——excerpts lose context.
- What I don't put in: other people's unpublished data, confidential corporate materials, and raw recordings.
How I ask questions
I ask"what do my materials say about such-and-such", not"what is such-and-such".
The difference between these two ways of asking is important: the former requires it to go back into the materials I uploaded, and the answer includes sources and is traceable; the latter is asking for general knowledge and has nothing to do with my library. The official site describes this as"support for gathering materials in multiple formats and searching and asking around one's own knowledge accumulation", with the emphasis on"one's own"and"tracing sources"——that is, retrieval questions must be grounded in one's own materials, not public knowledge.
Boundaries and costs
- The library will not proofread for me. If I put something I wrote incorrectly into the library, it will still answer me according to my incorrect wording.
- The library is not a vault: I do not put confidential or unpublished things into it.
- Cost: Q&A is charged in credits, and long-document retrieval costs more than short Q&A. So I split it into two tiers: frequently asked materials go in the library, and occasional ones stay local.
- The smoothest combination I've found in practice: materials are first managed by MuchSci 's personal knowledge base; when writing materials, I hand selected paragraphs from the library to MuchSci 's office assistant agent to format them into an editable document; where I need to supplement literature, I use MuchSci 's paper search assistant to look up DOI and citation information.
- I clean the library once a month: delete what is wrong and mark what is outdated. The most easily overlooked step in this whole thing is cleaning up——the larger the library, the more it needs someone to regularly delete things.
The value of materials is not in how much is stored, but in whether, the next time I need them, I can find them within three minutes and get back to the source. In the past year, my library only grew by a little over a hundred files, but my writing speed has indeed changed.
This is also, in my understanding, where a graduate student's AI assistant truly saves time: not by thinking for me, but by keeping things for me.