aerugo_

What I really want to know is - how do LLMs understand time series at all? Admittedly, even the best LLMs are not fantastic at analyzing time series and tabular data, but they also don’t completely suck at it. Why is that? They seem better at it that my intuition tells me they should be.

In my opinion we need a multi-modal model that is great at both tabular datasets and text analysis. Most analytical work in economics, policy, public health, medicine etc requires a combination of crosschecking between both. Current gen LLMs are not good enough at generating novel insights by looking at tables and text at the same time. I also haven’t have found any data on this so please serve it to be on a plate if I’m wrong.

Yeah absolutely. Once I am sure of what libraries I will need, I will probably remove the CDN dependencies and just bundle everything.

Ease of development mostly. I am still in the experimentation phase with this. Once I know what I want to use, I will bundle.

No, it's not "running" the file, it's the browser that runs it. For some reason, if I keep it as HTML, the file will not open in the browser when opened from a Sharepoint directory (mounted in the filesystem or opened through the browser), even though it runs perfectly fine when opened from a local folder. I discovered that simply renaming the HTML file to .aspx allowed it to be opened from Sharepoint. No idea why, I have zero knowledge of .NET and its ecosystem.

That's really neat! SimpleGantt needs to be a bit more feature-complete than is easily supportable with something like that, but I love the simplicity. I am an avid user of TiddlyWiki and this would be a great addon to that platform.

Interesting idea. I might do it, I see where you're coming from with it.

Great feedback! Thank you. Obvious now that you mention it. I just fixed it. Pull the new version and you will be able to add uncategorized tasks.

I love this idea, and will definitely do this.

Just FYI if anyone runs into this: I made a few updates to Prelude in the last few days. It now allows including only files tracked by git by using the -g option. You can also toggle case sensitivity of pattern matching (case insensitive by default) with the -c option.

New release https://github.com/aerugo/prelude/releases/tag/v0.3 and pushed to the Homebrew formula

Good idea, followed your advice for the next release.

Yes, the whole point of this tool is just to make it extremely simply to make a prompt out of a part of your repo, with some pattern matching on filenames. It's just 200 lines of bash, so obviously nothing fancier than that.

However, I just had a go at using Aider, and WOW - I have tried a bunch of these auto-commit tools before (GPT Engineer, Fume, etc) but I found them to be clunky and unintuitive. Aider has fixed a lot of these problems, and I absolutely love it. Can't believe I hadn't heard of it until now. Thanks! I see myself using this quite a bit, and using Prelude when I need to debug something where the Aider workflow gets stuck.

I was asked in a PM how I use the prompts. Basically, whenever I need to write some new code for a project, I start by making a prompt out of the relevant parts of the repo, paste it into GPT or Claude, and follow up with something like (random made up example):

——

I want to add a feature to my Content.js component. In the table that displays names of animals, I want to add an Edit button in a new column on each row. When clicked, it should bring up a modal that allows me to edit the name and emoji of the animal.

Review my codebase to see how CRUD actions are implemented for other features. Make sure you follow the established code patterns in the project.

Before starting to code, consider best practices for UX of a feature like this and make sure the implementation is user friendly and follows best practices.

After reviewing the code, describe how you will implement the feature make a plan for how to implement. After that, write the complete code for the implementation.

Interesting! Did you publish it? It’s a great idea to prioritize based on git history.

As for multiple stages, it means the tool itself is doing a few calls to the model. What do the code summaries look like? Just function and class doctrings? Getting the model to write summaries that are comprehensive enough to guide development but still more compact than the code itself seems like it may not be a trivial problem.

This is also neat, like simpler minimal version of Fume or GPT-Engineer. It seems really useful. Good job!

Prelude is rather meant for when you are pair-programming with a chat model and find yourself pasting a lot of context to remind the model of the state of your code. I always have terminals open anyway, so having it as a CLI tool rather than needing to make requests to a server, makes sense for me.

Just like this:

--- File: ~/GitRepos/kanot/prelude ---

Same here! Prelude respects .preludeignore, and also ignores annything in .gitignore.

Oh yeah, I hadn’t seen code2prompt - that’s indeed really similar. Prelude is a lot simpler though, just around 200 lines of bash in a single file, with only a single dependency (a CLI tool called tree). code2prompt, in contrast, requires you to install Rust and use Cargo.

Prelude is meant to really just do a single thing. I haven’t felt the need for things like templates, but I will give code2prompt a spin if I need more fancy things.

I was feeling bogged down by the repetitive copy-paste routine to build context prompts for debugging or generating code. This simple solution ended up saving me a lot of time, so I thought I'd share my workflow.

With the large context windows of OpenAI and Anthropic models, you can keep entire code repositories in context. But copy-pasting the file-tree and code from lots of places is annoying. Prelude is a tiny CLI tool that creates a context prompt from files in your directory and its subdirectories, then copies it to your clipboard or saves it to a text file. Prelude creates a prompt containing the file tree of your specified directory, minus the files you ignored in .gitignore and .preludeignore, followed by the content of each file. It adds a bit of context prompting at the end, and that's it. Ready to copy-paste into a new chat with your favorite AI-chat. Nothing more, nothing less. Here’s how it works:

Install Prelude with Homebrew:

brew tap aerugo/prelude brew install prelude

Simply run prelude in your working directory. Need files from a specific subdirectory? Use prelude -p dir. Want to specify a filename to save the prompt? Use prelude -f prompt.txt. Lots of files you don't want in the prompt? Put them in a .preludeignore file in the root folder where you run Prelude

Say that you know the names of the files you want to include, for example FooBar.py and BarFoo.md and they are somewhere in your file hierarchy.

Just run

prelude -M "foobar|barfoo"

and you will get a prompt with only those files. It also works with * wildcards.

Of course, it also works with the -P option, so you can also run something like

prelude -P backend/src -M "*.py"

To get all python files in the src directory of your backend.

PS. I'm sure this already exists. Lots of people must have done this. But nevertheless, I couldn't find a simple CLI tool to do just this, so I wrote one for myself, and then thought to share it.