Post

David Keyes
@dgkeyes.com

I just did a livestream focused on writing good prompts. I was planning to give Claude generic prompts like, “Make a bar chart on this data.” I hoped it would, by default, write some base R code, and then I could give specific instructions such as, “Use the tidyverse,” to have it write ggplot code.

David Keyes
@dgkeyes.com

But my experience was the total opposite. It used ggplot by default. This made me wonder if the models are now so good that you need to focus less on writing good prompts (and yes, I do think that using ggplot for plotting is better than using base R). What do you all think? #rstats

0 reposts2 likes

beefbacon.bsky.social
@beefbacon.bsky.social

Yeah I always have to remind myself that the models don’t think hierarchically, like how one teaches base R to understand foundations, then layers on package-relevant complexity. Also probably reflects volume of tidy vs base documentation/tutorials/blogs/etc

Nicola Rennie
@nrennie.bsky.social

I almost always use skills alongside prompts so things like which packages/formatting/structure I want isn't information I include directly in prompts anymore, so I can't really comment on prompting only

Nicola Rennie
@nrennie.bsky.social

But I think historically before using skills, most models returned ggplot2 code if asked for a chart (there are a lot more tutorials online about plotting with ggplot2 compared to base R I think, so I kind of assumed it was a bigger part of any training data).

David Keyes
@dgkeyes.com

Yeah, maybe ggplot isn't the best example. I also thought it might suggest base R code to read in and combine multiple CSV files, but it used {purrr}. I guess Claude likes the tidyverse!

Chraz
@edu.runlinea.com

Clearing global memory and CLAUDE.md only proves the tidyverse prior lives in the weights. If you want base R for a session, ban ggplot/purrr in the system prompt. Otherwise generic still means whatever dominated the training set.