How to use the Validate Data Analysis agent skill (validate-data)

What is the validate-data skill and what does it make the agent do?

An agent skill is a folder with a SKILL.md file that tells an AI agent how to handle a specific kind of task. This skill covers analysis quality assurance. Its description is: validate whether an analysis is accurate, well-supported, and ready to share or use for a decision.

Once installed, the agent inspects a report, notebook, spreadsheet, SQL query, dashboard or chart. It checks the question being answered, the methodology and filters, whether key numbers add up when recomputed, whether charts could mislead, and whether conclusions go beyond the evidence. It then lists required fixes.

• Checks the question, population, filters, metric definitions and baselines
• Recomputes the highest-impact numbers where it can
• Reviews charts for truncated axes, dual axes, inconsistent intervals and misleading titles
• Flags conclusions and causal wording the evidence does not support
• Separates blockers from caveats and records a confidence assessment

Install the validate-data skill from a ZIP file

Download the ZIP for the validate-data skill from this page. The package contains agents/openai.yaml and SKILL.md. Unzip it into one of these locations:

• Personal skills folder: ~/.claude/skills/ (available across your projects)
• Project skills folder: the .claude/skills/ folder inside a specific project
• Claude apps: upload the skill under Settings → Capabilities → Skills

Trigger the skill with a natural request

You do not need a special command. Describe what you want checked and point the agent at the artifact. The skill's description mentions reviewing methodology, calculations, comparisons, visuals, caveats, or conclusions, so requests using those words are the most likely to activate it.

Include the artifact itself, such as a file, notebook, spreadsheet, SQL query or dashboard link, because the skill can only verify what it can reach. You can also name the decision the analysis supports and the audience, since the skill checks whether the analysis fits them.

• "Validate this churn analysis before I send it to leadership. The file is churn_q3.ipynb."
• "Review the methodology and calculations in this spreadsheet before the pricing decision."
• "Check whether the conclusions in this dashboard go beyond the evidence."
• "Is this chart misleading? Flag blockers versus caveats."

Understand what the output looks like

The skill is built to produce a documented outcome rather than a general opinion. Expect the agent to separate verified findings from interpretation, label anything it could not verify, and state what would be needed to verify it.

Blockers are issues that materially affect the decision, such as an unreliable number, denominator, join, time window, population, comparison or conclusion. Caveats are minor issues that should not stop sharing. When the agent runs checks in a notebook or spreadsheet, the output is meant to include the artifact path or tab so the check can be reproduced.

• Blockers: material problems that should be fixed before sharing
• Caveats: minor issues or limits to disclose
• Unverified items: what could not be checked and what access would verify it
• Reproducibility references: file paths, query links, notebook paths or spreadsheet tabs used in checks

Know what the skill needs and what it does not do

The skill needs access to the analysis artifact for inspection. It can use optional connectors if your setup provides them: a structured data source connector, an operations logs connector and a spreadsheet workspace tool. Without these, it works from the artifacts you provide and labels what it cannot check.

The skill is not for profiling a raw dataset on its own. Dataset freshness and grain checks are handled by a companion skill, analyze-data-quality, which is not part of these files. The files also name product-business-analysis, jupyter-notebooks and visualize-data as companions. Their availability is not confirmed from these files, so check whether you have them before relying on them.

• Requires access to the analysis artifact (file, notebook, spreadsheet, SQL or dashboard link)
• Optional connectors: structured data source, operations logs, spreadsheet workspace
• Companion skills named: analyze-data-quality, product-business-analysis, jupyter-notebooks, visualize-data (availability not confirmed)
• Runs scripts: no
• Uses network: no

Safety notes on scripts and network access

According to the package, this skill does not run scripts and does not use the network. That means the agent's work is limited to reading and reasoning over the artifacts you give it, plus any connectors you have configured.

Review the connectors you enable before running the skill on sensitive data, since those connectors are what let the agent reach source systems. Only give it access to analyses you are comfortable having inspected.

• No script execution is stated in the package files
• No network use is stated in the package files
• Connector access is the main route to source data, so limit it to what you need

Tips for getting useful results

Give the agent the decision and audience up front. A check for an internal experiment readout and a check for a board deck apply different bars. Name the artifact precisely, and include the time window and filters if you know them.

If you want a recommendation after validation, ask for it as a separate step. The skill points to product-business-analysis for that. If the data itself may be stale or at the wrong grain, run the data quality check first.

• State the decision and audience in your request
• Share the exact artifact path or link
• Ask for blockers and caveats separately
• Run data quality checks before validating when source trust is in doubt