Code Review of an R or Python Script With Claude

Save the invented script below as check_expected.py. It reads two invented files, joins them and computes an expected count per unit. It is meant to contain two problems.

import pandas as pd
import numpy as np

exposure = pd.read_csv("synthetic_exposure.csv")
rates = pd.read_csv("synthetic_rates.csv")

merged = exposure.merge(rates, on="cell_code", how="left")
merged["expected_claims"] = merged["earned_units"] * merged["rate"]

print("Rows after join:", len(merged))
print("Units after join:", merged["earned_units"].sum())

sample = merged.sample(frac=0.5)
total_expected = sample["expected_claims"].sum()
total_units = sample["earned_units"].sum()

summary = sample.groupby("cell_code")["expected_claims"].sum()
print(summary)
print("Expected per unit:", total_expected / total_units)

What you should see: About 20 lines of Python. Line 7 is the merge. Line 13 is the sample. Troubleshooting: If you prefer R, ask Claude to translate the invented script first, then review the translation yourself before using it.

The two planted issues:

1. Line 7, join that can duplicate rows. If the rates file has two rows for the same cell_code, the left join returns an exposure row for each, so units and expected claims are counted twice.
2. Line 13, random step with no seed. The sample is different on each run, so results cannot be reproduced for a reviewer.

Other points may also be raised, such as unmatched codes giving missing rates after the join. That is a fair point, not a planted one.

What you should see: Your own notes naming the two lines. Troubleshooting: If you want the review to be a fair test, write the answers down before you upload.

1. Sign in to Claude and start a new chat. A regular chat is enough, with no project needed.
2. Click the "+" button at the lower left of the chat box and choose "Add files or photos". Or paste the script into the message.
3. Choose your file.

Claude's upload page lists PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB, JSON and XLSX as supported document types. A script file with a .py or .R extension is not on that list, so save a copy as check_expected.txt before uploading, or paste the code into the chat.

What you should see: The file shown as an attachment in the message box. Troubleshooting: If the upload is refused, check the file type and use the TXT copy.

Paste this prompt.

Review the attached script, which uses synthetic data. List risks under these headings: hard-coded paths, unhandled missing values, silent type coercion, joins that duplicate rows, and random steps with no seed. For each risk, name the line number, explain what could go wrong in one or two sentences, and give a small invented test case I can run to see it. If a heading has no risk, say so. Do not rewrite the whole script.

What you should see: A list under each heading, with line numbers. The merge on line 7 and the sample on line 13 should appear, and the hard-coded file names on lines 4 and 5 may appear too. Troubleshooting: If line numbers look off, ask Claude to quote the line. Different editors count blank lines differently, so check against your file.

Take the test case for the join risk. A small invented version is below.

Exposure file (3 rows):

Rates file (4 rows, with Cell 3 repeated):

Run the script on these files. The join produces 4 rows, not 3, and the units after the join are 900, not 600 (100 plus 200 plus 300 plus 300). That confirms the duplication.

For the seed risk, run the script twice and compare the printed summary. Without a seed the sample, and so the summary, can differ between runs.

What you should see: "Rows after join: 4" and "Units after join: 900". Troubleshooting: If your output differs, check that the CSV headers match the column names in the script.

1. Decide how to handle repeated rate rows. That is a decision about your data and belongs with you and your reviewer. Possible choices include removing the duplicate, or adding a second key column.
2. For the seed, set a fixed random_state value in the sample call.
3. Run the script again on the invented files and confirm that rows and units match the exposure file, and that two runs give the same summary.

What you should see: "Rows after join: 3" and "Units after join: 600". Troubleshooting: If you ask Claude for a fix, test it on the invented files first. Never accept a fix because it looks right.

Claude's list is a first read. The reviewer signs off on model code through your team's normal process.