Skill: Data Analyst

The Data Analyst skill turns Claude into a data scientist who writes clean, reproducible Python analysis code. Instead of one-off scripts that nobody can re-run, you get a structured workflow: load, inspect, clean, analyze, visualize — each step in its own code block with the reasoning written out alongside it.

What This Skill Does

When triggered, Claude follows a five-step workflow and outputs the Python code for each step:

Load — reads the data with pandas, handling different file formats (CSV, JSON, Excel, Parquet).
Inspect — runs head(), info(), describe() to understand shape, types, and distributions before touching anything.
Clean — handles missing values, fixes column types, removes duplicates, and documents every transformation.
Analyze — runs the actual analysis: groupby, pivot tables, correlations, statistical tests.
Visualize — produces charts with matplotlib or seaborn, labeled and ready to drop into a report.

Two hard rules keep the output useful: Claude always outputs the Python code block (not just a description), and it explains the statistical significance of findings — not just "the mean is 42" but whether the difference actually means something.

How to Install

Create the skills directory:
```
mkdir -p .claude/skills/data-analyst
```
Save the template below as .claude/skills/data-analyst/SKILL.md.
Trigger it with a prompt like: "Analyze sales.csv using the data-analyst skill. I want to know which regions grew fastest last quarter."

SKILL.md Template

# Data Analyst Skill

## Triggers
- "Analyze data"
- "Visualize this"
- "Clean this CSV"

## Workflow
1.  **Load:** Read data using pandas.
2.  **Inspect:** Check `head()`, `info()`, `describe()`.
3.  **Clean:** Handle missing values, fix types.
4.  **Analyze:** GroupBy, Pivot, Correlation.
5.  **Visualize:** Use matplotlib/seaborn.

## Rules
- Always output the python code block.
- Explain the statistical significance of findings.

Real-World Usage Example

Prompt:

Analyze this CSV using the data-analyst skill. The file is orders.csv with columns: order_id, region, date, amount. Which region had the highest average order value last quarter, and is the difference statistically significant?

Expected output:

1. Load

import pandas as pd

df = pd.read_csv('orders.csv', parse_dates=['date'])

2. Inspect

print(df.head())
print(df.info())
print(df.describe())

Check for nulls in amount and confirm date parsed correctly.

3. Clean

df = df.dropna(subset=['amount', 'region', 'date'])
df['amount'] = pd.to_numeric(df['amount'], errors='coerce')
df = df.dropna(subset=['amount'])

4. Analyze

last_quarter = df[df['date'] >= '2024-10-01']
aov = last_quarter.groupby('region')['amount'].mean().sort_values(ascending=False)
print(aov)

from scipy import stats
top_region = aov.index[0]
second_region = aov.index[1]
t, p = stats.ttest_ind(
    last_quarter[last_quarter['region'] == top_region]['amount'],
    last_quarter[last_quarter['region'] == second_region]['amount'],
    equal_var=False
)
print(f"p-value: {p:.4f}")

5. Visualize

import matplotlib.pyplot as plt
import seaborn as sns

sns.barplot(data=last_quarter, x='region', y='amount', estimator='mean')
plt.title('Average Order Value by Region (Last Quarter)')
plt.ylabel('Average Amount ($)')
plt.xticks(rotation=45)
plt.tight_layout()
plt.savefig('aov_by_region.png', dpi=150)

Findings: The top region's average order value is higher than the second region's. The p-value tells you whether that gap is likely real or just sampling noise. A p-value below 0.05 means the difference is statistically significant; above 0.05 means you can't rule out random variation with this sample size.

Customization Tips

Pin your stack. If you use polars instead of pandas, or plotly instead of matplotlib, swap the library names in the workflow section. Claude will use what you specify.
Add a "no silent failures" rule. Append "Raise on unexpected null counts. Print a warning if more than 5% of rows are dropped during cleaning." This surfaces data quality issues instead of hiding them.
Require a data dictionary. Add "Before cleaning, output a table of columns, types, and null counts." You get an audit trail of what the raw data looked like.
Set a chart style. Add "Use sns.set_theme(style='whitegrid') before plotting." Your charts will look consistent across analyses.

Combining With Other Skills

technical-writer — After the analysis is done, hand the findings to technical-writer to produce a structured report. The analyst provides the numbers; the writer provides the narrative.
code-reviewer — Run code-reviewer on the generated analysis script to catch issues like chained indexing (df[df['x']>0]['y'] = ...) that pandas silently ignores.
security-auditor — If the dataset contains PII, run security-auditor on the script to check for data leakage in logs or saved charts.

Common Mistakes to Avoid

Not specifying the file format. "Analyze this data" without mentioning CSV vs. JSON vs. Excel leads to wrong loader code. Always name the file and format.
Skipping the inspect step. If Claude jumps straight to analysis without checking types and nulls, the results can be wrong in ways that don't error — strings where numbers should be, silent NaN propagation.
Ignoring sample size. A mean difference with 12 data points is not the same as one with 12,000. The skill asks for statistical significance for this reason — read the p-value, don't just look at the bar chart.
Not saving charts to files. plt.show() doesn't work in headless environments. The template uses plt.savefig() — keep it that way unless you're running in a notebook.

Share this article