How to use the Analyze Data Quality agent skill

What are agent skills and what does this one do?

An agent skill is a folder containing a SKILL.md file that gives an AI agent a defined workflow for one kind of task. When this skill is installed and you ask the agent to judge whether data is reliable, it follows a structured audit instead of giving a quick opinion.

The workflow starts by clarifying what the dataset represents, its intended unit of analysis (grain), and the downstream use. It then builds a compact profile covering row count, column names and types, candidate keys, duplicate rates, min and max timestamps, null rates and distinct counts. It runs core checks for completeness, uniqueness, validity, consistency, integrity, timeliness and volume, and looks for changes over time such as new releases or backfills.

• Reports concrete evidence, the analytical risk, likely causes, and the smallest useful fix or automated test.
• Adapts checks to the data shape: event data, dimension tables, fact tables, ML feature tables, or experiment data.
• Prefers putting the SQL or Python behind its findings in a companion notebook so a person can inspect the code.

Install the skill from the ZIP download

Download the ZIP for the analyze-data-quality skill from the page where you found this guide. The skill's files are agents/openai.yaml and SKILL.md.

To install it for yourself across projects, unzip the download into ~/.claude/skills/. To install it for a single project, unzip it into that project's .claude/skills/ folder.

If you use the Claude apps instead, upload the skill under Settings → Capabilities → Skills.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into the project's .claude/skills/ folder
• Claude apps: Settings → Capabilities → Skills, then upload

How do you trigger the skill with a natural request?

You do not need to type the skill name. The skill's description says it applies when the task is to check data quality, reconcile conflicting sources or metric definitions, or decide whether evidence is safe to cite. Ask in those terms and the agent should pick it up.

Be specific about the table, dashboard or query and what you plan to do with it. Mentioning a change such as a release, migration or a date when numbers shifted helps the agent prioritise temporal checks.

• "Is the orders table trustworthy enough to cite in our Q3 board pack?"
• "Revenue dropped on 14 March after the migration. Check the events table for duplicates or missing partitions."
• "Reconcile these two sources for active users and tell me which metric definition to trust."
• "Before we train on this feature table, check for leakage and null-rate spikes."

What do you need before running it?

Querying tables requires access to a SQL or data source through the referenced connectors (~~structured_data, and ~~operations_logs for freshness and lineage checks). Code-based checks need Python or a Jupyter environment for the companion notebook.

The skill runs scripts and uses the network. Confirm your permissions before pointing it at live production tables.

• Access to a SQL or data source via ~~structured_data if querying tables
• ~~operations_logs if you want freshness and lineage checks
• Python or Jupyter if code checks are run

Safety notes: scripts, network and production data

This skill runs scripts and uses the network, and its workflow may run queries against live tables through connectors. Review what the agent intends to run and check that the account it uses has read permissions only for the data you intend it to see.

For sensitive production data, review the companion notebook and queries before accepting the findings. The skill is not designed for designing KPIs or for reviewing the conclusions of a finished report; use design-kpis or validate-data for those.

• Check permissions before the agent queries live tables.
• Read the companion notebook before relying on its results.
• Use design-kpis for defining metrics and validate-data for QA of reports or charts.

Tips for getting useful results

Give the grain and the expected primary key if you know them. Many apparent quality problems turn out to be mixed-grain data, partial backfills, late-arriving rows or duplicated joins, and the skill confirms grain before interpreting anomalies.

Provide a baseline such as a prior week, a prior schema or a trusted reference table. Rates compared over time are more useful than raw counts.

• State the intended unit of analysis, such as one row per order.
• Name a trusted reference table to reconcile against.
• Ask for the companion notebook so you can rerun the checks yourself.