Data Viz

Create publication-quality data visualizations from tabular data files. Uses qsv to profile and prepare data, then generates Python charts with best practices for clarity, accuracy, and design.

Cowork note: If relative paths don't resolve, call mcp__qsv__qsv_get_working_dir and mcp__qsv__qsv_set_working_dir to sync the working directory.

Steps

1. Understand the Request

Determine:

• Data source: CSV/TSV/Excel file, or query results from a prior analysis
• Chart type: Explicitly requested or needs to be recommended
• Purpose: Exploration, presentation, report, Data Schematic component
• Audience: Technical team, executives, external stakeholders

2. Profile the Data with qsv

a. Index and detect: Run mcp__qsv__qsv_index, then mcp__qsv__qsv_sniff to detect format and encoding.

b. Understand structure: Run mcp__qsv__qsv_headers and mcp__qsv__qsv_count to get column names and row count.

c. Profile columns: Run mcp__qsv__qsv_stats with cardinality: true, stats_jsonl: true to understand types, ranges, and distributions. Read .stats.csv to inform chart design:

• type → choose appropriate axis type (numeric, categorical, date)
• min/max → set axis ranges
• cardinality → determine if column is categorical (low) or continuous (high)
• nullcount → note missing data that could affect the chart

d. Check distributions: Run mcp__qsv__qsv_frequency with limit: 20 on columns you plan to plot — this reveals the actual values and whether grouping or filtering is needed.

e. Run moarstats for visualization hints: Run mcp__qsv__qsv_moarstats with advanced: true. Read the enriched .stats.csv for chart design decisions:

f. Preview data: Run mcp__qsv__qsv_slice with len: 5 to see actual values and formats.

3. Prepare the Data

Use qsv to prepare visualization-ready data:

• Filter: mcp__qsv__qsv_search or mcp__qsv__qsv_sqlp to subset rows
• Aggregate: mcp__qsv__qsv_sqlp for GROUP BY, window functions, computed columns
• Select columns: mcp__qsv__qsv_select to keep only what's needed
• Sort: mcp__qsv__qsv_sqlp with ORDER BY for ordered categories or time series

Export the prepared data to a CSV file for Python to read.

4. Select Chart Type

If the user didn't specify, recommend based on data and question:

5. Generate the Visualization

Write Python code using matplotlib + seaborn (default) or plotly (if interactive requested):

import matplotlib.pyplot as plt
import seaborn as sns
import pandas as pd

# Load the prepared CSV
df = pd.read_csv('prepared_data.csv')

# Set professional style
plt.style.use('seaborn-v0_8-whitegrid')
sns.set_palette("husl")

# Create figure with appropriate size
fig, ax = plt.subplots(figsize=(10, 6))

# [chart-specific code]

# Always include:
ax.set_title('Clear, Descriptive Title', fontsize=14, fontweight='bold')
ax.set_xlabel('X-Axis Label', fontsize=11)
ax.set_ylabel('Y-Axis Label', fontsize=11)

# Format numbers appropriately
# - Percentages: '45.2%' not '0.452'
# - Currency: '$1.2M' not '1200000'
# - Large numbers: '2.3K' or '1.5M' not '2300' or '1500000'

# Remove chart junk
ax.spines['top'].set_visible(False)
ax.spines['right'].set_visible(False)

plt.tight_layout()
plt.savefig('chart_name.png', dpi=150, bbox_inches='tight')
plt.show()

6. Apply Design Best Practices

Color:

• Use a consistent, colorblind-friendly palette
• Highlight the key data point or trend with a contrasting color
• Grey out less important reference data

Typography:

• Descriptive title that states the insight, not just the metric (e.g., "Revenue grew 23% YoY" not "Revenue by Month")
• Readable axis labels (not rotated 90 degrees if avoidable)
• Data labels on key points when they add clarity

Layout:

• Appropriate whitespace and margins
• Legend placement that doesn't obscure data
• Sort categories by value (not alphabetically) unless there's a natural order

Accuracy:

• Y-axis starts at zero for bar charts
• No misleading axis breaks without clear notation
• Consistent scales when comparing panels
• Appropriate precision (don't show 10 decimal places)

7. Save and Present

1. Save the chart as PNG with descriptive name
2. Display the chart to the user
3. Provide the Python code so they can modify it
4. Suggest variations (different chart type, different grouping, zoomed time range)

Data Preparation Recipes

Time Series

qsv_sqlp: SELECT date_col, SUM(value) as total
  FROM data GROUP BY date_col ORDER BY date_col

Top-N Categories

qsv_frequency: --select category_col --limit 10

Or for aggregated values:

qsv_sqlp: SELECT category, SUM(amount) as total
  FROM data GROUP BY category ORDER BY total DESC LIMIT 10

Distribution

qsv_stats: Check min, max, mean, stddev, cardinality
qsv_moarstats: --advanced for kurtosis, bimodality
qsv_sqlp: SELECT FLOOR(value/10)*10 as bin, COUNT(*) as cnt
  FROM data GROUP BY bin ORDER BY bin

Correlation

qsv_select: Pick the two numeric columns
qsv_stats: Verify both are numeric types with reasonable ranges

Comparison Across Groups

qsv_sqlp: SELECT group_col, AVG(metric) as avg_metric, COUNT(*) as n
  FROM data GROUP BY group_col ORDER BY avg_metric DESC

Notes

• Always profile with qsv first — stats and frequency reveal the right chart type and catch data issues before plotting
• For large files, use mcp__qsv__qsv_sqlp to aggregate before passing to Python — don't load millions of rows into pandas
• If interactive charts are requested (hover, zoom), use plotly instead of matplotlib
• Specify "presentation" for larger fonts and higher contrast
• Multiple charts can be created at once (e.g., "create a 2x2 grid")
• Charts are saved to the current directory as PNG files
• For data quality issues discovered during profiling, recommend /data-clean before visualizing