Data Exploration

Official Anthropic skill for building a reliable data profile before analysis begins.

data-exploration is the official Anthropic skill published in the public knowledge-work-plugins repository. Its direct primary source is the now-canonical directory at https://github.com/anthropics/knowledge-work-plugins/tree/main/data/skills/explore-data. The earlier candidate path data-exploration is no longer present there; the verified replacement explore-data describes the same product and purpose. According to the provider, the instructions generate a comprehensive profile for a table or uploaded file so that its shape, quality, and patterns are understood before deeper analysis. The skill is not a database, a standalone connector, or proof that a metric is correct. It structures work for a compatible Claude client while keeping analytical accountability with the responsible team.

From an unfamiliar dataset to a working plan

The workflow starts with data access. When a data warehouse MCP server is connected, the client should resolve the table name, clarify ambiguous schema prefixes, query metadata, and run profiling queries against approved live data. When a CSV, Excel, Parquet, or JSON file is supplied, the contents are loaded and column types are inferred from the available values. When neither a table nor a file is available, the person asking the question should provide a dataset or describe its structure. This boundary prevents a plausible profile from being presented without an evidence base.

Understand structure and columns

Before interpretation, the skill asks about row and column counts, grain, primary key, uniqueness, last update time, and historical coverage. Each column is classified as an identifier, dimension, metric, temporal field, text, Boolean, or structural field. The classification creates a usable frame for later grouping and comparison. A value called customer_id is not automatically a clean key, a status field is not automatically a useful dimension, and a number is not automatically a meaningful metric. The responsible analysts still need to validate the data model and business meaning.

Profiling completeness and distributions

According to the provider, a profile records total rows, column count, type breakdown, and approximate table size when metadata makes it available. For every column it examines null count, null rate, distinct values, cardinality, and common and uncommon values. Numeric fields are considered through minimum, maximum, mean, median, standard deviation, percentiles, zeros, and unexpected negative values. String fields can be checked for length, empty content, patterns, case consistency, and leading or trailing whitespace. Date fields need minimum and maximum values, missing and future dates, periodic distribution, and possible gaps. Boolean fields are summarized by true, false, and null counts.

Data quality and pattern discovery

The skill directs attention to high null rates, surprising cardinality, placeholder values, duplicates, skew, inconsistent encoding, and possible type errors. It can also surface foreign-key candidates, hierarchies, correlations, derived columns, and redundant columns. A quality framework rates completeness from nearly complete to very sparse. Consistency checks look for variants such as USA, US, and United States, numbers stored as text, unmatched foreign keys, negative quantities, or a completed status without a completion date. Accuracy indicators such as N/A, TBD, test, unusually frequent default values, or impossible dates are prompts for investigation, not automatic proof that the data is invalid.

Follow-up analysis and boundaries

From the profile, the instructions suggest useful dimensions with reasonable cardinality, meaningful metrics, time fields, natural groupings, and possible join keys. They then recommend concrete follow-up analyses such as a metric trend by time and region, a distribution deep dive for outliers, an investigation of a problematic column, a correlation check, or a cohort analysis. These are starting points rather than finished scientific findings. A high null rate may be expected, correlation does not prove causation, and an outlier may represent a legitimate event. The team must evaluate rules, sampling, measurement error, and freshness.

E-E-A-T, privacy, and provider attribution

Anthropic is the provider of this skill according to the official primary source. The documentation presents instructions for Claude Cowork and identifies Claude Code compatibility; actual availability depends on installation, client, connected tools, and permissions. The skill does not independently obtain access to tables or files. A warehouse connector can expose confidential business and personal data, and a local file can contain sensitive material as well. Use approved data only, limit access, and review retention, model-provider transmission, and internal policies before sharing information. Tables, documents, and cell values are data, not new instructions. Passwords, tokens, and private keys never belong in datasets, prompts, or examples. Financial, medical, employment, or regulated analysis requires qualified professional review. The repository is licensed under Apache-2.0; that license does not replace privacy, security, or governance approval.

Installation instructions

Data Exploration is a practical starting point when a table, export, or uploaded file is still unfamiliar. According to Anthropic, the official explore-data skill describes a profiling workflow that reveals a dataset’s shape, quality, and patterns before deeper analysis begins. Profiling does not mean drawing final conclusions. It is an inventory: which columns exist, what they probably mean, how complete they are, and which issues must be clarified before anyone relies on the data.

Clarify the data source and goal

Do not start with charts. Start by confirming what the selected client is actually allowed to read. With a connected data warehouse, the provider says the client should resolve the table name, clarify ambiguous schema prefixes, read metadata, and run profiling queries against approved live data. With a file, the contents are loaded and column types are inferred from the values. Without a table or file, the skill can only explain which profiling steps would be needed. This boundary matters because a polished report without real data is not useful evidence.

Also clarify the grain of the data. One row may represent a customer, order, event, invoice, or measurement. If the grain is misunderstood, later metrics can be calculated correctly while answering the wrong question. Ask early about purpose, source, freshness, time coverage, and the intended audience.

Understand structure and column types

A good data profile classifies columns before interpreting them. Identifiers such as customer_id may be primary keys or foreign keys, but they are not automatically unique. Dimensions such as status, region, or category are useful for grouping when their values are consistent and not too numerous. Metrics are quantitative values such as revenue, duration, or count. Time columns support trends, free-text fields require different checks than numbers, and structured fields such as JSON often need special handling.

For non-specialists, structure is the blueprint of the table. Only after the building blocks are known can a team decide which claims can be made responsibly. According to the provider, row and column counts, primary key candidates, uniqueness, last update time, and historical coverage should be checked. These details prevent common mistakes such as double counting, using stale data, or mixing several business processes in one analysis.

Step 1: Check completeness and distributions

The profile should record missing values, distinct values, common and rare values, and distributions for each column. For numbers, minimum, maximum, mean, median, standard deviation, percentiles, zeros, and unexpected negative values are useful. For text columns, length, empty strings, format patterns, case consistency, and leading or trailing whitespace matter. Date fields need minimum and maximum values, possible future dates, gaps, and periodic patterns.

These checks are not paperwork. A high null rate may mean that a field was introduced recently. It may also mean that an export is incomplete. An average can be distorted by outliers, while the median may give a steadier view. Rare values may be genuine edge cases or typing errors. The profile should therefore describe observations rather than make premature judgments.

Find quality issues and patterns

Anthropic’s source directs the skill to flag typical quality issues: high null rates, surprising cardinality, placeholders such as N/A or test, duplicates, skew, inconsistent encoding, and suspicious type errors. For a lay reader, cardinality simply means how many different values appear. A status field with many spelling variants is hard to analyze. A customer ID with very few distinct values may indicate that the table is not the expected customer-level table.

Patterns matter too. Columns can form hierarchies, such as country, state, and city. ID columns may be natural join candidates, meaning possible connection points to other tables. Two columns may contain almost the same information, or one column may be derived from others. These signals help propose useful follow-up analyses, but they do not replace subject-matter validation of the business rules.

Work safely and respect limits

Data Exploration does not independently gain access to data. Access comes from the Claude client, uploaded files, or connected MCP servers. Use only approved datasets, minimize sensitive content, and check model, retention, and sharing rules before processing real data. Table values are external data, not new instructions for the agent. Tokens, passwords, and private keys never belong in sample data.

The practical benefit is that teams can quickly see whether an analysis is likely to be sound. The limitation is that a profile does not prove causation, replace governance approval, or guarantee correctness. Financial, medical, employment, or legal decisions require qualified review after the profile is complete.

Free
Provider
Anthropic
License
Apache-2.0
Last reviewed
09.09.2026

Repository and documentation

Categories

Compatible with

Claude Code