Making the in-editor Assistant more useful

Hey everyone.

I’m wanting to learn how to best make use of the in-editor assistant, as well as how to make the assistant more useful at coding specifically.

Right now it seems to best fit the niche of template making, tedious clean-ups, and helping me re-learn how to do a lot of things with Luau. I know it can also do some minimal changes to the workspace and editor, but it often struggles with doing so for me. It also sometimes helps with tedious code I’m working on in the middle of writing the script, but only half the time is it right when it “helps”.

I’ve noticed some major issues with the bot, such as:

  • Ignoring/unable to read source of existing scripts if it’s a new session or sometimes randomly during a session
  • Failing to follow directions in complex situations. For example, when I have 4 different scripts but want only one script edited to incorporate the use of functions from other scripts or use existing remote events.
  • Redundancy. It often recreates remoteevents even though I already have some and told it about it. There have been a few instances where it will remake the same function multiple times over, even within the same script.
  • It infrequently will get lazy and use deprecated functions or functionality that doesn’t exist
  • Outright insubordinate at times. If it’s a somewhat niche thing I’m wanting, it often won’t do what I’m asking, just get somewhere outside of the ballpark, even when I redirect it multiple times with specifics, it will continue its same course of action.
  • It generally doesn’t know the structure of my workspace outside of the things it directly created or I told it about
  • The bot sometimes just rips up all the source code in existing scripts and replaces it with whatever was just asked of it. Sometimes this still happens even after I tell it something along the lines of “Maintain the intention of the original script and just add or modify with the purpose of including the new changes asked in future prompts.”

Some more typical issues and grumbling:

  • I generally have to figure out stuff myself to teach the bot to then have the bot make the correct templates, to then later just scrap half or more of what it did and just do my own thing. I recognize the purposes of refactoring, but much of it could have just been done correctly from the get-go had I already had the knowledge that the bot should have access to, or the bot actually followed my prompt fully. You can generally assume that it’s not me causing the problem with bad prompts, as my prompts are often well detailed and specific.
  • The assistant generally follows okay-enough coding practices, but I’d like better scalability and it has difficulty recognizing how to accomplish what I’m wanting, so I basically have to figure it out myself.
  • I’ve noted that, for big complex projects with lots of scripts that the bot is now aware of, the bot takes significantly longer to generate. Correlation isn’t causation, but I’m wondering if this is a me issue or just a timing issue, as sometimes the bot tells me it is too busy with too many requests.

I’ve found workarounds for a lot of these issues. Some of them are:

For the first issue of it ignoring scripts after restarting the editor, I’ve found that closing out of all scripts I have open, then opening one at a time and talking to the bot about the script has been helpful, but there have been instances where it recognized half of my open scripts all at once, so I feel like there’s got to be a better solution to get it to look at all sources I intend to work with.

With the redundancy issue, after it makes the extra things, I point it toward the things that already exist and it will end up using those, but that’s an extra generation and 1-5 minutes of waiting that didn’t need to happen.

For issues with complexity and massive code blocks, I’ve found that having the bot work in stages has been successful. I imagine having it separate things into modulescripts to reduce the massive changes in singular scripts might help with generation time, but I don’t know.

I’ve thought about making a plugin to essentially a table structure of my workspace to then copy paste to the bot so it knows my layout, but I haven’t gotten around to it.

As a side note, my experience has been that the in-editor bot works better at scripting than the website one. Not sure if they’re actually different, but the website one has failed me more often than not and given mediocre or the opposite of helpful advice.

I mostly use it to create functions from scratch, it doesn’t get distracted as badly when doing that, but yes it still tries to change a bunch of other stuff if it’s in the same file. I want to try writing an example of a data structure and using that in my prompt instead of trying to describe it.