yprincedec / data-quality-csv-auditor

name: data-quality-csv-auditor
description: Audits CSV datasets for missing values, duplicate rows, inconsistent data types, invalid values, and potential outliers. Use when a data analyst needs a reliable data-quality report and safe cleaning recommendations.
license: MIT
metadata:
llmskillhub:
version: 0.2.0
categories:

• coding

keywords:

• csv
• data-quality
• data-cleaning
• data-analysis
• python

capabilities:
network: false
filesystem: read
shell: false
secrets: []
compatibility: "Works with CSV files and text-based dataset descriptions. Requires access to the provided file for direct inspection."

Data Quality CSV Auditor

Purpose

Help students and data analysts identify data quality problems before performing exploratory data analysis or machine learning. Produce transparent findings, practical recommendations, and reproducible validation steps.

When to use

Use this skill when a user wants to inspect a CSV dataset, find missing values, detect duplicate records, investigate inconsistent formatting, review possible outliers, or prepare a dataset for analysis.

Workflow

1. Understand the data

• Confirm the dataset's purpose and intended use.
• Inspect available column names, row counts, data types, and a small sample.
• Ask questions when column meanings or valid ranges are unclear.
• Never claim to have opened or inspected a file that is unavailable.

2. Check completeness

• Count missing values in each column.
• Calculate missing-value percentages when row counts are available.
• Identify columns with unusually high missingness.
• Recommend appropriate handling methods without automatically filling missing values.

3. Detect duplicates

• Identify exact duplicate rows when data access permits.
• Check possible duplicate identifiers.
• Explain the difference between exact duplicates and records that merely share some values.
• Do not remove records without user approval.

4. Check consistency and validity

• Review inconsistent data types and formats.
• Identify leading or trailing whitespace and inconsistent category labels.
• Check dates, numeric ranges, and units against known rules.
• Distinguish confirmed invalid values from values that simply look unusual.

5. Investigate potential outliers

• Use suitable summaries, such as minimum, maximum, median, and quartiles, when data is available.
• Explain the method used to flag unusual observations.
• Do not assume that every outlier is an error.
• Request domain context when necessary.

6. Recommend cleaning actions

For each finding, report:

• The issue and affected column.
• The evidence and number of affected records, if known.
• A recommended action.
• The reason for the recommendation.
• Risks, assumptions, or decisions requiring user approval.

Preserve the original dataset. Recommend creating a separate cleaned copy and documenting all transformations.

7. Validate the cleaned data

After authorized cleaning, compare the original and cleaned datasets. Recheck row counts, missing values, duplicate rows, data types, and relevant constraints. Report remaining issues and never claim a test passed unless it was actually performed.

Output format

Provide a concise report with these sections:

1. Dataset overview
2. Missing-value summary
3. Duplicate-record findings
4. Consistency and validity issues
5. Potential outliers
6. Prioritized cleaning recommendations
7. Validation checklist
8. Limitations and assumptions

Use tables where helpful. Calculate statistics only from accessible data.

Safety and reliability

• Do not invent dataset statistics or test results.
• Do not disclose confidential or personal data unnecessarily.
• Do not send dataset contents to external services.
• Do not overwrite or delete original data without authorization.
• Explain uncertainty and ask clarifying questions.
• If execution tools are unavailable, provide suggested code or instructions without claiming they were run.

Example

A user provides a sales CSV with missing prices and duplicate order records.

The skill should report the affected columns and counts if available, explain that duplicate order IDs need business-rule verification, recommend reviewing missing prices before analysis, and provide validation steps. It must not invent counts or automatically delete records.

Expected outcome

A reproducible, understandable data-quality assessment that helps users prepare reliable datasets for exploratory data analysis and machine learning.