such a keen way of levelling up your knowledge base - i keep notes in obsidian which is automatically in markdown, and now i wanna try opening it thru cursor and play around w what i can do!
What kinds of notes/non-code content would you want an LLM to help you with?
I use Cursor for my personal website and it has a very annoying habit of injecting or even substituting slop copy over my own copy, it keeps trying to "help" me with my plaintext and I wish it would stop!
Can only imagine it would be even worse with a simple markdown file like I use in Obsisian
All of the different cursors are different type-of-guys (or gals, depending on the base model), and we have to be specific about which guy we're talking about.
The main little-guys cursor has are:
- auto-complete guy
- command-K guy (aka inline-editing-via-prompt guy)
- command-L guy (aka "ask mode")
- command-I guy (aka "agent mode")
Auto-complete guy always uses the cursor-fast model (i think), which is a custom LLM trained by cursor to quickly autocomplete code. You can (and should) turn off autocomplete on files that contain copy.
For the latter 3 cursor modes, you get to choose the model and the prompt yourself, and then all the regular rules of prompt-engineering and model selection apply. Personally, I like using gemini-2.5-pro and claude-3.7-sonnet thinking for most things these days.
Once you have a good prompt-engineering system, there isn't much that I wouldn't want an LLM to help me with, at least in Ask Mode, where it doesn't even change the copy or the note, we just get to talk back and forth about stuff, it gets to search the internet, etc...
thank you for writing this so i don't have to! i have become more invested in my obsidian setup because of this. i do experience some tension between wanting to use cursor and being hesitant about potentially uploading all my files to an LLM provider, but i'm getting over that - especially for work-related writing.
i keep thinking about this and how this is based off comparison w notion, whereas my mind wondered and used this thesis in my database storage. that is, your essay title convinced me to store my "world model of a human" not in jsonl, but in a simple txt/md like string. i used to have a schema json-ed of a few components that would force each created world model to comply with schema. now, bc i no longer have the jsonl, the txt/md string of world can be more open-ended / schema-less.
i mean, almost all web apps are already specified in plain text and stored on Github.
but I think I see what you mean. As in, how long until web apps are essentially just "prompts" that compile into malleable HTML or React apps, which let us interact with our (also plaintext, hopefully) data on the fly?
Geoffrey Litt is a researcher of this emerging idea in end-user programming:
such a keen way of levelling up your knowledge base - i keep notes in obsidian which is automatically in markdown, and now i wanna try opening it thru cursor and play around w what i can do!
What kinds of notes/non-code content would you want an LLM to help you with?
I use Cursor for my personal website and it has a very annoying habit of injecting or even substituting slop copy over my own copy, it keeps trying to "help" me with my plaintext and I wish it would stop!
Can only imagine it would be even worse with a simple markdown file like I use in Obsisian
All of the different cursors are different type-of-guys (or gals, depending on the base model), and we have to be specific about which guy we're talking about.
The main little-guys cursor has are:
- auto-complete guy
- command-K guy (aka inline-editing-via-prompt guy)
- command-L guy (aka "ask mode")
- command-I guy (aka "agent mode")
Auto-complete guy always uses the cursor-fast model (i think), which is a custom LLM trained by cursor to quickly autocomplete code. You can (and should) turn off autocomplete on files that contain copy.
For the latter 3 cursor modes, you get to choose the model and the prompt yourself, and then all the regular rules of prompt-engineering and model selection apply. Personally, I like using gemini-2.5-pro and claude-3.7-sonnet thinking for most things these days.
Once you have a good prompt-engineering system, there isn't much that I wouldn't want an LLM to help me with, at least in Ask Mode, where it doesn't even change the copy or the note, we just get to talk back and forth about stuff, it gets to search the internet, etc...
Oohh cool thank you, I hadn't even looked into the most basic customization/realized it existed
np :)
thank you for writing this so i don't have to! i have become more invested in my obsidian setup because of this. i do experience some tension between wanting to use cursor and being hesitant about potentially uploading all my files to an LLM provider, but i'm getting over that - especially for work-related writing.
i keep thinking about this and how this is based off comparison w notion, whereas my mind wondered and used this thesis in my database storage. that is, your essay title convinced me to store my "world model of a human" not in jsonl, but in a simple txt/md like string. i used to have a schema json-ed of a few components that would force each created world model to comply with schema. now, bc i no longer have the jsonl, the txt/md string of world can be more open-ended / schema-less.
This is so well written
How long until our web apps are specified in plain text, and then an LLM-backed interpreter outputs an application on the fly?
i mean, almost all web apps are already specified in plain text and stored on Github.
but I think I see what you mean. As in, how long until web apps are essentially just "prompts" that compile into malleable HTML or React apps, which let us interact with our (also plaintext, hopefully) data on the fly?
Geoffrey Litt is a researcher of this emerging idea in end-user programming:
https://www.geoffreylitt.com/2023/03/25/llm-end-user-programming
Also, Claude Artifacts and v0 are already getting somewhat scarily close.