CSV Anomaly Detector

Scans a CSV sample for data-quality anomalies — outliers, missing values, type drift, ragged rows, duplicates, and categorical inconsistencies — and returns a structured findings report.

This is an InTouch skill: a SKILL.md prompt document the InTouch AI assistant loads and follows. It is not code and runs no binaries.

What it does

Given a sample of CSV text, the skill profiles every column and surfaces problems a downstream load or analysis would trip over:

• Missing values — per-column blank/null/NA counts and percentages, with high-severity flags above 20% missing.
• Type drift — values that don't match a column's dominant inferred type (text in a numeric column, malformed dates), with example line numbers.
• Numeric outliers — mean/median/min/max/stddev per numeric column, flagging values beyond 3σ or outside the IQR fence.
• Ragged rows — rows whose field count differs from the header.
• Categorical anomalies — rare categories and casing/whitespace variants of the same value.
• Duplicates — fully duplicated rows and duplicated key values.
• Domain rules — any plain-language rules you supply (e.g. "amount must be positive") are checked and violations reported.

Each finding gets a severity (HIGH / MEDIUM / LOW) and a recommended fix.

What to feed it

• A CSV sample as plain text — header row plus data rows. The first N rows of a larger file are fine; the skill reports findings as sample-based and notes truncation.
• Optionally, a line or two of domain rules after the CSV describing expected constraints.

What it returns

A Markdown report with:

• A header line (rows scanned, column count, truncation flag)
• A Summary of top findings, highest severity first
• A Column Profile table (column, inferred type, missing %, distinct count, notes)
• A numbered Findings list with severity, affected column, concrete example values + line numbers
• Recommended Actions for each high/medium finding

If the sample is clean, it says so explicitly and still returns the column profile.

Files

• SKILL.md — the skill prompt (frontmatter + instructions)
• manifest.yaml — InTouch Hub metadata
• README.md — this file

Provider: Blue Isle · License: MIT