Gemini Gem as a Code Explainer

1. Go to gemini.google.com and sign in.
2. Open the sidebar. Google's help page shows the path as Open Sidebar, then Gems.
3. Look for the option to create a new Gem. The label may read New Gem, and the exact words can differ between accounts.

What you should see: A form with fields for a name, a description and instructions. Troubleshooting: If you cannot find Gems, your account may not include them, or your admin may have turned them off. Ask IT.

1. Type a name such as Code explainer.
2. If there is a description field, type: Explains pasted Stata, R or Python code in plain terms. Never asks for data.

What you should see: The name and description filled in. Troubleshooting: Keep the name short so you can find it in the sidebar later.

Paste the block below into the instructions field.

You explain code for an applied economist. The user pastes a block of Stata, R or Python code with no data.

For every block of code the user pastes, do these things in order:
1. Name the language and say in one sentence what the block does overall.
2. Walk through the code in blocks of one to five lines. For each block, say in plain terms what the lines do, one block at a time.
3. Give the code back with a short comment above each block. Do not change any line of code.
4. Add a section called "Lines that may not behave as described". List each line where the behavior could differ from what a reader would assume, or from what a comment says. Examples: a lag function that takes the previous row and not the previous year, a merge that can duplicate rows, a missing value that is dropped without notice, integer division, a sort that changes row order. For each item, say what to check, and do not say the code is wrong unless the code itself shows it.
5. End with the line "Explanation only. Run the code on a synthetic file and read every line before relying on it."

Never do any of these:
- Never ask for data, and tell the user not to paste any if they do.
- Never choose, change, recommend or defend a statistical or econometric model.
- Never state a result, a coefficient or a number that the code would produce.
- Never claim you ran the code.
- If part of the code is unclear, say what is unclear and ask the user what the line is meant to do.

1. Save the Gem using the button on the form (often labelled Save or Create).
2. Open the Gem from the sidebar to start a chat with it.

What you should see: An empty chat with your Gem's name at the top. Troubleshooting: If the Gem does not follow the rules in the first reply, check that the instructions saved in full.

Paste the snippet below. It is invented, and the file named in it does not exist.

* practice_panel.dta is invented
use practice_panel, clear
xtset firm_id year
gen growth = sales / L.sales - 1
summarize growth

What you should see: A plain-terms walk through five short blocks. The Gem should say that xtset firm_id year declares a panel with firm_id as the unit and year as the time variable. It should also say that L.sales is the value of sales one period earlier in the time variable, so a firm with no row for the previous year gets a missing value for growth, and that summarize growth summarizes the non-missing values. Troubleshooting: If the Gem says L.sales is "the previous row", correct it in the chat. Stata's lag operator works on the time variable set by xtset, not on row position.

Paste this invented R snippet next.

library(dplyr)
panel <- read.csv("practice_panel.csv")
panel <- panel %>%
  group_by(firm_id) %>%
  arrange(year) %>%
  mutate(growth = sales / lag(sales) - 1)

What you should see: A walk through the read, the grouping, the sort and the growth line. In "Lines that may not behave as described", the Gem should flag lag(sales): in dplyr it takes the previous row within the group, not the previous year. If a firm has a gap year, the growth for the row after the gap is computed against the earlier year and looks like a one-year growth. Troubleshooting: If the Gem misses the lag point, ask: "Does lag() here check that the previous row is exactly one year earlier?"

You can see the difference between the two snippets yourself.

1. Create a small synthetic file with one firm and three rows: year 2021 with sales 100, year 2022 with sales 110 and year 2024 with sales 132. There is no 2023 row.
2. Run the Stata snippet on it in Stata. The growth for 2022 is 0.10, and for 2024 it is missing.
3. Run the R snippet on the same file in R. The growth for 2022 is 0.10, and for 2024 it is 0.20, because 132 divided by 110, less 1, is 0.20, computed against the 2022 row.
4. Add your own year check to the R code if gaps matter for your data, for example by comparing each row's year with the previous row's year plus one.

What you should see: Two different answers for 2024 on the same file. The Gem's flag was right, and the R code needs a check. Troubleshooting: If your results differ, print the data after sorting and look at the year column row by row.

1. Copy the Gem's commented code into your own file only after you have read it line by line.
2. Delete any comment that you cannot defend.
3. Run the code in your firm's own environment, never inside Gemini.

What you should see: Your file with comments you have checked. Troubleshooting: If a comment is wrong, fix it in your file and tell the Gem in the chat, so you can see whether it corrects its explanation.