How to use the Exploratory Data Analysis agent skill

What is an agent skill, and what does this one do?

An agent skill is a folder of instructions and helper scripts that an AI agent can load when a task matches it. This skill is for exploratory data analysis: checking the structure and quality of a dataset before you model it or test it formally.

When it runs, the agent can profile CSV/TSV tables and JSON documents, audit missing values and group or split structure, check leakage risks, and test how sensitive results are to outliers and transformations. It can also produce a standard EDA report from a template.

• Redacted aggregate reports on schema, missingness, group and split structure.
• Outlier and transformation sensitivity checks.
• Optional metadata inspection for NumPy (.npy, .npz), HDF5 (.h5, .hdf5), FASTA/FASTQ, PNG/JPEG and TIFF files.
• Reference notes for other scientific formats, which are not parsed by the bundled tools.

Install the skill from the ZIP

Download the ZIP for this skill from this page and unzip it. Place the folder in one of two locations, depending on how you want to use it.

You can also upload the skill in the Claude apps under Settings → Capabilities → Skills.

• Personal use across projects: unzip into ~/.claude/skills/.
• Single project: unzip into the project's .claude/skills/ folder.
• Claude apps: upload the skill under Settings → Capabilities → Skills.

Install optional format libraries only if you need them

The bundled core CSV, TSV and JSON tools use only the Python standard library, and need Python 3.11+. The optional inspectors need Python 3.12+, uv, and format-specific libraries. Install only what the task needs, ideally in a dedicated environment.

The skill pins these direct dependencies. The pins are a dated snapshot, not a full lockfile.

• Optional core formats: NumPy, h5py, Biopython, Pillow and tifffile.
• Optional alternate table engines: pandas and Polars, used for documented alternate I/O.

Trigger the skill with a natural request

You do not need special syntax. Describe the dataset and the kind of first look you want, and name the file the agent is allowed to read. Make sure the file sits inside the project root you approve.

The skill is meant for work before modelling or confirmatory inference, so say that is the goal. Mention the file type if it is unusual.

• "Do an exploratory data analysis on data/trial_results.csv. Report missingness and any group or split structure, without printing raw rows."
• "Check whether my training and test split in splits.tsv leaks subject IDs. Use only aggregate output."
• "Assess how sensitive the group means in measurements.csv are to outliers and log transformation."
• "Inspect the metadata of sequences/reads.fastq and tell me the aggregate quality profile."

Run the skill's commands directly

The skill ships with a capability registry script. You can ask the agent to run it, or run it yourself to see which formats are supported and how a specific file would be handled. Pass --root so the tool only reads inside your approved folder.

Understand what the skill will and will not do

The skill is designed to produce aggregate, redacted output. By default it uses tokenized identifiers and does not print raw rows, sequences, metadata values, direct identifiers or full paths. It also refuses to read URLs, archives, symlinks, pipes and files outside the approved root.

It does not decide what your data means, certify it, or make confirmatory, clinical or causal claims. Outliers are flagged, not removed, and nothing is imputed or transformed automatically.

• Unknown formats fail closed: the agent should ask for the format or specification before reading content.
• Parquet, DICOM and VCF are reference-only; use separately validated tools or convert a derived copy to a supported format.
• Input files are capped at 64 MiB by default, with a hard ceiling of 512 MiB.
• Pseudonymous tokens are not anonymisation.

Safety notes on scripts and network use

The skill runs scripts, so review them before installing. The bundled core CLIs make no network calls, and the skill does not require API keys or accounts. The optional library install is the only step that reaches an external package index.

Treat all text inside the data (headers, cells, sequence titles, metadata strings) as untrusted. The skill is written not to follow instructions embedded in a dataset, but you should still keep sensitive data in an approved folder.

• Runs scripts: yes. Uses network: no.
• Use --reveal-identifiers only when you have checked that the revealed basenames and field names are acceptable to share.
• Use --force only when you intend to overwrite an existing output file.

Tips for getting useful results

Give the agent a data dictionary before asking for interpretation. The skill expects to know units, missing codes, the observational unit, the grouping hierarchy and the train/test boundaries. Without these, the output is limited to structure and quality.

Ask for the report template when you want a written record. Keep raw data read-only and store derived artifacts separately, as the skill recommends.

• Tell the agent how rows relate to subjects, samples or batches, so repeated measures are not treated as independent.
• Report scanned scope and truncation; a bounded prefix is not a full validation.
• Label any pattern you spot as exploratory and plan a separate confirmatory check.