Join the conversation as we shape the future of AI in RStudio!
This space is for our private beta participants to share their real-world experiences with the Posit Assistant and Next-Edit Suggestions. Ask questions, showcase how you’re using the new tools, and help us refine as we continue to iterate. Your feedback directly impacts the tools we're building for the entire data science community.
First impression is amazing. It will be a great tool and addition to Rstudio. Much needed for people who don't want to leave Rstudio for other AI IDEs. Best thing about it is that it understands the analysis flow and other things that a professional statistician does. Great job team.
Definitely enjoying the beta, however I've found myself moving back and forth from other AI coding tools for the following reasons:
I have a tough time getting the assistant to consistently put code in a file. Often times, I will ask it to write me a function to do some task, and it will write it but execute the function in the console. This means I can't examine the code. What I really want is something like claude code that will edit a file and allow me to accept/reject the changes.
Speaking of changes, would be great to see them in the file, like in cursor.
Lastly, the lack of a plan mode is a huge pain point. I want the AI to take on larger changes and refactors, but I don't want to give it free reign. Plan mode in cursor and claude code has been such a boon to my workflow that I would love to see this in Rstudio as well.
First, might be good to understand the use case for plan mode. When I know what I need to do and how to do it , plan mode is really good. I can give Claude enough detail without having to write the code myself. The benefit of plan mode is that I can see, from a high level, what claude is going to do and then either accept the plan or correct claude. This strikes a nice balance between verifiability in what will be produced and how much effort I need to give in the prompting.
As an example, our unit testing structure requires generating sql queries and then comparing them against a static string. If left to its own devices, claude will typically just check that some substrings exist in the generated query rather than use my preferred method of testing. But if I use plan mode, and tell claude Use function_compare() to compare the generated query against the expected query. Look at file_1.go, file_2.go to see examples of how this is done. mimic this as closely as possible, then claude is more likely to succeed. This is good because I get the outcome I want without investing too much time into instructing claude or writing out the code myself. Usually, I will even ask claude to provide code snippets of how it plans to implement something in the plan document which increases my trust that I can succeed in whatever task I am setting out to do.
The plan also gives the AI some "rails". The plan is something the I or the AI can always refer back to, and I have found that if I have a good plan but the AI goes off the rails, I can always stop, edit the plan, and then ask the AI to implement the plan from scratch.
NES produces a mess.
I turned on this feature, however, quickly found it messy. First of all, it over-thinks my next steps and offers something unexpected. At least, it has offered me the relevant code in 20% cases. All the other was useless. It could be just written off and given up, but the issue is that it gets stuck in a strange mode when any arrow moves your cursor to the right, and the mouse is never able to put the cursor in the right position. Esc doesn't help. Ctrl+■■■■+; helps, but it is annoying to click this (new) combination too frequently.
I would recomment to re-design this feature to make it somehow softer, or avoidable.
A note about Claude. It looks like the prompt you generate for Claude requests to provide a detailed answer. I find it excessive. Every time, it does too much analytics and generate conclusions I don't need. He assumes knowing my problem better than I do, therefore offers me unnessary solutions I haven't asked for.
In general, the tool is very helpful, however, I would propose to introduce a keword to select the detalisation level required. For example:
/1 what is the right name of line width parameter in ggplot?
/2 how to reorganize data in dt, so every row contains one observation?
/3 I want to build Shiny app where I can choose the country, economic indicator and draw the timeline of its development at the screen, all using API to IMF. Create it for me.
Only one word required as the answer with keword /1
A couple of rows of code in response to /2
Complete code or analysis in response to question with keyword /3
I find the AI integration in RStudio extremely useful. However, a pressing problem has recently emerged. When the agents attempts to make changes to the quarto code, it constantly requests permission, and the ‘always allow’ command does not work.
Thanks, Alexei. I shared your feedback with the developers and they mentioned several streams of work that should improve both the underlying model and UI in the coming weeks. I'll circle back here when those ship.
Regarding the constant permission requests, can you try again with the newest Posit Assistant (0.0.93)? We've overhauled the permissions system in the last day or so, hopefully it was fixed--I was not able to repro with this version, at least.
If I create a new R script, drop some rows of code there to inspire AI (giving context on where data is located and some formats), then ask a question about next steps -- it actually doesn't have access to Untitled1 code pane. It tries finding it in the RStudio buffer, in the console output, but fails and starts inventing its own code.
Checking back in here. The thinking effort and planning mode are a big improvement and the assistant feels very fast...noticeably faster than claude code (under some models). Is that intentional or did I imagine that?
One more piece of feedback. Sometimes, I just want to open up Rstudio and try something quickly in an Untitled and unsaved .R script. I might: open Rstudio, hit Cmd+N to open a new R script, and then just hack away at something only to close Rstudio and not mind that I will never get that file back again.
The assistant would be great for this kind of work, because I'm looking to get something quick done and just get feedback as opposed to persisting the code. I can copy the code from the assistant, and if I have a file open it looks like I can insert the code at the cursor, but it would be awesome if i could prompt the assistant and have it create a file I could then save if I wanted to. At the very least, this would make inspecting the code and making required edits a much smoother experience.
General suggestion.
The window of RStudio is now divided into 3 vertical sections. Assistant, code, var/plots panels. Each of them became too tiny to fit all the information, even if I use my 22" screen. Of course, you can propose to buy 40" screen. However, my suggestion is to have the multi-window app, so I can place panels on other side-screens with a lower size/resolution. This partially works for the Plot panel (zooming it out, put on the other screen), but... there are no left/right, clean buttons on it, thus it is not fully-functional.
There is the option to move the Posit Assistant to a different quadrant and not use the Sidebar. For example (you have to scroll the list of options, it's below the fold):
IMPORTANT: I just found a bug where this isn't sticking and it goes back to the Sidebar after a restart, at least on macOS. Sigh, will open an issue and get this fixed ASAP!
Better than before. However...
At least for me, what panels are the most frequently used:
Source (in some cases source should be split too - when I do non-exact copy of previous code)
Console
Environment
Plots (this one fortunately can be detached, unfortunately without buttons)
and additionally -
Help
Viewer
Files
Now, Assistant Tool appears and goes to the position 2. This actually means I should keep 5 (FIVE) panels permanatly visible, plus one panel for additional Viewer/Help/Files. It becomes too difficult to sort them out over 4 available panels.
Also, my eyes are not perfect, thus I use big size fonts.
It's not a claim, just some reasoning. Can be void for other users.
Just read the release announcement. Is there no ability to provide an API key from anthropic so I don't have to pay the $20/month? I like posit AI, but it doesn't make sense for me to pay for claude code and posit Ai
