Write prompts that get good R code

I went live to work out how to write prompts that get good R code, and the audio broke at the start, so the first stretch of the recording is mostly a loop. Once I got going, I tried to get a generic, non-tidyverse answer from Claude and kept getting the tidyverse code I'd have written myself, which I hadn't expected.

You'll learn:

• How to set up a clean test environment (a new project, a blank CLAUDE.md, an incognito chat) so your files, memory, and global instructions don't shape the R code the AI writes
• Why a generic question gets a generic answer and a prompt that names the approach gets more specific code
• How project files, global instructions, and skills can quietly steer the R style an AI uses, such as tidyverse over base R
• A way to compare how different Claude models handle the same R prompt by switching models and re-running the request
• How to run AI-generated chart code step by step to see whether it produces the bar chart you asked for

Tools mentioned: Claude, Claude Code, Solo, Claude Opus, Claude Sonnet 4.6, arrow, tidyverse, ggplot2, purrr, fs

Key moments:

• 00:00 Going live and audio trouble
• 15:18 Starting a new deportation data script
• 15:54 Asking how to read in Parquet files
• 19:14 Running the first AI-written chart code
• 20:42 Creating a clean project to avoid polluted context
• 25:06 Prompting for a bar chart of arrests by month
• 27:08 Testing the same prompt in an incognito Claude window
• 31:50 Splitting the data into monthly CSV files
• 36:13 Claude unexpectedly loads a data tidying skill
• 41:49 Splitting files under 4 MB for regular Claude
• 45:46 Switching between Claude models
• 50:48 Takeaway: be as specific as possible
• 51:31 Wrapping up and next livestream