Anthropic Data Analysis Skills

Anthropic shipped an official skill pack for data analysis. MIT licensed. Markdown-based. Covers the full lifecycle from data loading through final report. It is the first first-party data-science skill bundle from a frontier lab.If you have ever pasted a CSV into Claude and watched it write reasonable but inconsistent analysis code, this is the structured replacement.What Actually ShipsThe repo is organized around the standard data-analysis workflow:Load and clean — pandas idioms for reading CSVs, JSON, Parquet, and SQL. Encoding detection, dtype inference, and structured handling for missing values, duplicates, and dirty timestamps.Exploratory data analysis — a structured EDA skill that produces a consistent profiling report on any DataFrame: dtype audit, distribution summaries, correlation matrices, outlier flags, and inferred categorical groupings.Statistical testing — proper hypothesis framing skills for t-tests, chi-square, ANOVA, Mann-Whitney, and Kruskal-Wallis. Each skill includes assumption checks, effect-size calculation, and multiple-comparison correction (Bonferroni, Holm, Benjamini-Hochberg).Visualization — matplotlib, seaborn, and plotly recipes with built-in chart-type selection logic. The skill knows when a histogram beats a density plot and when small multiples beat a single panel.Reporting — Jupyter-native output skills that produce notebook-ready tables, charts, and Markdown summaries.Every skill is a Markdown file with frontmatter metadata, an input contract describing required arguments, an output contract describing what the skill returns, and reference code the agent uses as a template.Why This MattersUntil now, agent-driven data analysis was a hit-or-miss exercise. The agent would write reasonable code on a good day and bizarrely inconsistent code on a bad day. There was no structured library of patterns the agent could rely on.This skill pack is the canonical answer. Anthropic published the patterns they actually use internally for data work. The statistical testing skills are the most valuable in the pack — proper hypothesis framing with assumption checks and multiple-comparison correction is exactly the work that AI agents tend to skip.How to Use ItClone the repo. Point your agent runtime at the skills directory. Claude Code, Cursor, Windsurf, Nanobot, and any agentskills.io-compliant runtime auto-discover the skills as available tools. From a Jupyter notebook, you can invoke a skill directly: ask the agent to 'run the EDA skill on df' and the agent loads the skill, executes the structured workflow, and returns notebook-ready output.For team deployments, pin a specific commit so the skill pack is reproducible across analyses. Fork the repo to add company-specific skills (your event taxonomy, your KPI definitions) on the same format.Where It Falls ShortTwo real limitations. First, the pack is general-purpose — it does not know your data domain. Healthcare, finance, and ad-tech analyses each need domain-specific skills layered on top. Anthropic published the foundation, not the domain layer.Second, the visualization skills are conservative — solid choices, not creative ones. If you need polished publication graphics, plan for a designer to take the skill output and refine it. The skill produces good first drafts, not finished deliverables.Who Should Install ItInstall it if you run Claude or any compatible agent on data work and you want consistent, reproducible analysis output. That includes data scientists, analytics engineers, product teams running A/B tests, and any company building an internal data copilot.Skip it only if your agent work is unrelated to data analysis. The skill pack is deep but focused — it is a data analyst, not a general-purpose knowledge pack.Related ResourcesArticle: GPT-5.5 just launched — the new model you can pair with these skills for stronger statistical reasoning.Tool: Gemini Enterprise Agents — load this skill pack into a Gemini agent for hosted data work.Repo: Microsoft Magentic-One — wire the skills into the Coder specialist for structured analysis runs.MCP server: Anthropic Claude Code MCP — pair these skills with Claude Code's file and bash tools for end-to-end notebook automation.