What are you doing right before you open your LLM?

Knowledge management
1

I’ve noticed something about my own workflow that I can’t quite shake.

Most of the time when I reach for a LLM on a research task, I’ve already been doing something else for a few minutes first — searching my vault, pulling up a few notes, trying to assemble enough context to even formulate a decent question.

By the time I open the LLM, I’ve already done a surprising amount of work just to get ready to ask.

I’m curious whether this resonates with others here, or whether I’m just doing something inefficient.

What were you doing in the 5–10 minutes before you opened your LLM the last time you used it for research or knowledge work?

I’m not asking about the task itself — more the ritual right before. What were you looking at, pulling together, or trying to remember?

1 Like
2

I’ll answer my own question since it’s only fair.
My ritual has two modes depending on what I’m trying to do.

Writing mode (I know what I’m working on): I open Obsidian and search for the note that holds my current argument structure, usually a note I titled something like “product-v3” or “essay-wyz.” I copy that. Then I go find the last thing I actually wrote (usually buried in a daily note) and copy the last few paragraphs. I paste both into Claude. That takes about 5–7 minutes, and it’s mostly just navigating — I usually know what I’m looking for, I just have to find it.

Exploration mode (I’m trying to think through something I don’t fully understand yet): I do almost nothing. I open the LLM and re-explain my whole project in the first message from memory. The problem is that my vault has three years of notes in it, and some of the best synthesis I’ve ever done is in there — things I’ve completely forgotten I figured out. But I can’t get that synthesis into the session without already knowing where it is, which defeats the purpose. So I start from scratch, and the LLM answers from its training, not from my three years of reading on the topic.

The part that bothers me most is the failed search. Sometimes I’ll spend 10 minutes in Obsidian trying to locate a specific note I’m almost certain I wrote — running three different search queries, following backlinks, opening dead ends — before giving up and just explaining the concept to the LLM from scratch. Which is kind of absurd. I wrote that note precisely so I wouldn’t have to re-derive the idea. But the retrieval cost is high enough that sometimes it’s faster to just… start over.

Curious if anyone else has the same split — and whether you’ve found anything that makes the exploration-mode session actually use what you’ve already built.

1 Like
3

How are you naming your notes?

1 Like
4

For the exploration mode issue - I’ve found a project-context.md per project really helps. Keep it to 5-10 lines and update it each session. Paste it first when you open Claude - it skips the whole re-explain. What worked for me: name notes as outcomes or questions rather than topics - that makes the search problem much easier.

2 Likes
5

My note-taking is a mnemonic system that helps me become more intelligent and productive. Text Analysis has been more valuable than LLM.

When I create new notes, the filename is {ZETTELKASTEN-KEY} {KEYWORD PHRASE}. The Zettelkasten key is YYYYMMDDhhmm. The “Keyword Phrase” represents the theme or note context.

The key never changes, but as the note evolves, the keyword phrase might. That’s because the keyword phrase telegraphs what is in the file.

When I’m searching for something in Obsidian, I tend to find something topical. I may then use the Graph Analysis plugin’s Jaccard tab to see what might be related to that text.

Note-taking is block-oriented, maps of content, atomic notes, and then specific deliverables. Old notes get deleted.

What I haven’t been able to do is get OpenClaw or Claude to consistently index my Obsidian vault, then cite what it finds.

I use LLM for developing topical clusters, sentiment analysis, and challenging ideas, not idea or concept generation.

Both, even with different models, cite non-existent notes. When they cite the correct note, the content doesn’t match what is stated.

If I want to find something, I go back to Omnisearch, Graph Analysis, or RegEx searches. Or I start updating a MOC on the topic to improve context in my notes.

Everything about note-taking is to get the meaningful part of the note in my meat sack of a brain.

There are times when SystemSculpt can take a pile of notes, list similarities in concepts, or create an outline for a specific audience.

Unless you have an in-house LLM, I’m confident you’re training someone else’s product rather than creating deliverables for your own personal benefit.

What makes it easy to find answers or specific notes is the vault structure. The interlinking of concepts, extractions, and observation creates context.

I’m learning Python text analysis to do things that the Graph Analysis plugin cannot. And taking notes on primary research feels more productive than conversing with an LLM.

OpenClaw used 6.5 million tokens over four hours. The session kept introducing an AI bias, wouldn’t consistently challenge ideas, and kept offering to write the guide.

I was able to get to my objectives faster, digesting notes into indexes, then reassembling them into articles or lists of questions. Then revisiting sources to flesh out the guide.

6

It depends on what the note is used for.

One of my activity is collecting feedback from users. In that case, the name will always start with a date (YYYYMMDD) followed by the name of the product.

If am writing summaries of academic articles, all these notes start with summary followed by author name then title of the article. It the title is too long I will pick the most significant keywords.

I also have notes to organize myself: processes, index, etc. This is a bit ad-lib right now.

I try to be consistent and coherent with the naming as much as possible.

@CawlinTeffid , any reason you were asking?

7

Thanks for sharing @hittjw

There is a lot to unpack here.

I have never tried OpenClaw but I am using Claude a lot. I am also using Claude Code which I think is way better than Claude for browsing the makrdown notes.

The main challenge with any LLM is that there is a limit to how much it can ingest before it becomes stupid. Behaviors like making stuff up or not being able to provide the exact source of a piece of information are symptoms that the LLM was fed too much information for its little digital brain.

One of the problems I am trying to solve is how to go around that limit. The idea would be to start with a small note (like a MOC) that would provide the right direction based on the task I have given it.

I don’t like having long conversation with LLMs but one thing I find very useful is to have them generate ideas that I wouldn’t think of by myself. And sometimes, just like when you are brainstorming with other people, it is not the ideas of the LLM that are interesting, it what new ideas they trigger for me.

I am still very early in the process but I would like to have a AI-based tool that would assist me in my knowledge work.

8

I asked because I mostly find notes by name, so the problem you described sounded like it might stem from unhelpful note names.

9

I haven’t gone down the path of local LLM. Instead, I’ve content-aware indexed the vault (a copy of it). This way I do not have to ingest into a local llm. I indexed 500 notes at a time, each in a different claude session; this avoids usage timeouts; weekends are our friends for usage.

In another pass, using the index, I found every occurance of properties and intenal wiki links. These are housed in a sqlite db.

the last study I did was for a junior highschool teacher who was interested in analyzing student stories for their grammatical and style content. Search patterns we were looking for in a deterministic way was valuable; we stored those in json files to guide the AI for preset prose.

But the coolest thing was letting the AI loose and telling it to analyze the writing styles for patterns that were not deterministic. When we compared the two resulting json files, the creativity factor was much higher for those writers who did not follow preset, if you will, passages relating to character arcs, themes, and plot developments. These writers also had higher abilities with grammar and syntax.

The teacher and I are going to work on a syntax highlighting mechanism and a web site app that will allow students to write short stories and highlight grammar, syntax, and style choices in near real time with explanations about their choices. The AI is very good, even with smaller models, at identifying text patterns.

We’ve started making markdown files to store methologies behind the given json files. In our case we also stored results in csv files to analyze via spreadshet.

This indexing makes querying incredibly fast and flexible. I’m not totally sold on the use of LLM locally; it just appears too deterministic. But that could be my ignorance.

It’s fast in this space right now… keep reading and learning and experimenting.

1 Like