Working in your editor
Programming with LLMs for Data Practitioners
Julia Silge | Posit PBC
Hello!
Day 2: Build with AI
- Intro to AI coding assistants
- Working in your editor (this session!)
- Steering and extending
- Agentic workflows
Lowest rung, highest mileage
- Today we live in your editor, with chat, inline completions, and quick fixes
- This is the most frequent, lowest-autonomy way to work
- You stay in the driver’s seat the whole time
Writing with AI
Completions appear as you type
- The assistant suggests the next line or block, greyed out, as you type
- Accept with Tab, or keep typing to ignore it
- It reads the surrounding code and your session for context
Steering with comments
- Write a comment describing what you want, then start the code
- For example:
# crashes per county, sortedand begin a dplyr pipeline
- Clear names and comments lead to better suggestions
Accept, edit, reject
- Read before you accept! A suggestion is not an answer
- Accept and then edit to fit, or reject and rephrase
- Completions are low-autonomy, so control stays with you
Habits that go wrong
- Accepting code you have not read
- Letting a suggestion pull you off what you meant to do
- Trusting plausible-looking code on data you have not checked
Two kinds of models
- Completions fire on every keystroke, so they must be fast and cheap
- They use small, low-latency models tuned to predict code at your cursor
- Chat and agents use larger, more capable, general purpose models that are slower and cost more
- The trade-off you feel: instant guesses versus slower, thoughtful help
Not every provider offers completions
- That small completion model is special, and not every provider ships one
- Anthropic, the Claude API we have been using, has no suitable completion model
- So with the workshop Anthropic key you get chat and agents, but not inline completions
- To experience completions today, you need Posit AI (can use the free trial) or GitHub Copilot
- Chat and agentic work, where we spend most of today, are the same either way
Completions
Demo
Follow along if you have Posit AI or Copilot!
In a script with the crash data loaded:
- Start typing
crashes <- read_and watch completion finish the path - Write a comment for a small task (“count crashes by city”) and let it suggest the code
- Accept, edit, reject: notice how it feels to stay in control
Explore your data
The assistant already knows your session
- Yesterday’s lesson, now in practice: it can see your data frames, columns, and plots
- So describe what you want, and skip pasting the data or the schema
- Ask questions in plain language and get back code that runs on your data
Exploring the crash data
Demo
In chat, ask Posit Assistant:
- “Plot the number of crashes by hour of day”
- Run it, look at the result, then ask for one change (labels, color, a different geom)
Your turn
Activity
Build a visualization of the crash data with chat or inline completions:
- Try something like crashes by county, by day of week, or severity by a risk factor (DUI, night, teen driver)
- Iterate! Ask for a change like sorting, labels, or a different chart
- Notice that it works from the data already in your session
- Share what you made with your neighbors
15:00
Debug and fix
When something breaks
- The fastest everyday win is to paste nothing, just ask
- Run your code, hit the error, and ask “what went wrong and how do I fix it?”
- The assistant can see the error, your code, and your session all at once
Your turn
Activity
Open 15-debug-crashes and run it from the top:
- When it errors, do not fix it yourself! Ask Posit Assistant “what went wrong?”
- Apply the fix, run again, and repeat
- There is more than one bug, so keep going until it runs clean
15:00
Now you take the wheel
I need a volunteer!
Demo
One of you drives on the big screen:
- Hit one of the bugs live and work with the assistant to fix it
- The rest of us will coach you
- What context does it need?
- Do we believe the fix?
Thanks!
Illustrations from unDraw