Analyze a Dataset With Pandas Step-by-Step

This prompt makes the model write commented, step-by-step pandas code for a real exploratory analysis on your DataFrame — profiling, null handling, IQR outlier detection, and then answering a specific business question. Reach for it when you have a dataset loaded and want a defensible analysis pipeline rather than a wall of unexplained code. Listing your actual columns and types is what lets it pick the right operations instead of guessing at your schema.

What you can expect back

# 1. EDA — understand shape, types, and where data is missing
print(df.shape)
print(df.dtypes)
print(df.isnull().sum())
print(df.describe(include='all'))

# 2. Handle missing values — drop rows missing the key fields, flag amount nulls
df = df.dropna(subset=['region', 'refunded'])
df['amount'] = df['amount'].fillna(df['amount'].median())  # median resists outliers

# 3. Outliers in amount via IQR — summarize rather than auto-delete
q1, q3 = df['amount'].quantile([0.25, 0.75])
iqr = q3 - q1
low, high = q1 - 1.5*iqr, q3 + 1.5*iqr
outliers = df[(df['amount'] < low) | (df['amount'] > high)]
print(f"{len(outliers)} outlier orders ({len(outliers)/len(df):.1%})")

# 4. Refund rate by region
rates = df.groupby('region')['refunded'].mean().sort_values(ascending=False)
print(rates)
# Meaningfulness: compare top region's rate to overall and check sample size per region
print(df['region'].value_counts())

Illustrative example — your results will vary by tool and inputs.

Get sharper results

• 01Get the exact dtypes right (especially datetime vs. object and category vs. string) because the code branches on them — a date stored as text will break any time-based step.
• 02Tell it your row count order of magnitude; for millions of rows it should suggest vectorized or sampled approaches instead of operations that quietly blow up memory.
• 03Specify whether outliers should be removed, capped, or just reported, since auto-deleting them can silently distort the very business answer you're after.
• 04Ask for the statistical test by name (e.g. a chi-square or two-proportion z-test) when the question is about whether a difference is 'meaningful,' so it quantifies significance instead of eyeballing it.
• 05Request that it print intermediate results after each block so you can sanity-check the data as the pipeline runs rather than trusting the final number blind.

Adapt it for your case