Claude Code for Data Scientists: 35+ Agent Skills That Replace Your Notebook Boilerplate

Every data scientist has the same first 50 lines. Import pandas. Load the CSV. Check dtypes. Count nulls. Describe the distributions. Plot the histograms. It is the same boilerplate in every notebook, every project, every dataset. Agent Skills encode that entire opening ritual - plus your team's specific domain knowledge, data dictionaries, and analysis patterns - into a folder Claude loads once and executes automatically.

I cataloged 35+ data science skills across 9 repos. The ecosystem is deeper than I expected. K-Dense-AI has 144 scientific skills (RDKit, Scanpy, BioPython - the real stuff). Anthropic's official sql-queries skill covers 5 SQL dialects. The official OpenAI jupyter-notebook skill has 2.4K installs. Streamlit published their own agent skill. There is even a publication-quality matplotlib skill with journal presets for Nature, Science, and Cell.

Here is what the data-analyst skill produces compared to a generic prompt. Without a skill, you ask "analyze this CSV" and get a wall of df.describe() output. With the skill installed, every analysis follows the What / So What / Now What framework:

What: Monthly churn rate spiked from 3.2% to 5.8% in April, driven entirely by the Basic plan tier (8.1% churn vs 1.9% for Pro).

So What: At this rate, Basic plan revenue drops $47K/month within 90 days. The Pro plan is healthy - this is a tier-specific problem, not a product-wide one.

Now What: Investigate what changed for Basic users in March (pricing change? feature removal? onboarding regression?). Run a cohort analysis on Basic signups from Q1 vs Q4 to isolate the variable.

Same data. The skill forces structure: finding, business impact, recommended action. That is the difference between "here are some numbers" and an analysis a stakeholder can act on.

---

Why Do Agent Skills Beat ChatGPT Uploads for Data Work?

The default "upload a CSV and ask questions" workflow has three problems:

1. No memory. Every session starts fresh. You re-explain your schema, your business rules, your column naming conventions every time.
2. No scripts. The model writes code but cannot execute persistent analysis scripts. Your EDA pipeline, your feature engineering transforms, your model evaluation harness - they live in notebooks, not in the model's toolkit.
3. No domain context. "Revenue" in your dataset might mean ARR or MRR or gross revenue or net revenue. The model guesses. A skill encodes the answer.

Agent Skills solve all three. The SKILL.md carries your data dictionary, analysis patterns, and domain-specific gotchas. The scripts/ directory carries your pandas pipelines, plotting utilities, and model evaluation code. The references/ directory holds your schema documentation and statistical methodology notes. The agent loads it all when a data task triggers the skill.

---

What Data Science Skills Should You Install? (by Workflow Stage)

Stage 1: Exploration & EDA

Skill	What It Does	Source	Install
exploratory-data-analysis	Automated EDA across 200+ scientific file formats. Detects file type, generates markdown reports with quality metrics, statistical summaries, missing value analysis. Includes eda_analyzer.py script.	K-Dense-AI/scientific-agent-skills	npx skills add K-Dense-AI/scientific-agent-skills
analytics-data-analysis	Expert pandas/numpy/seaborn workflow: vectorized ops, descriptive stats, correlation, pivot tables, visualization templates. 551 installs.	skills.sh	npx skills add mindrally/skills --skill analytics-data-analysis
data-analyst	Senior analyst workflow: frame business questions as testable hypotheses, write/validate SQL with CTEs, cohort/funnel/hypothesis testing, What/So What/Now What reporting. 368 installs.	skills.sh	npx skills add borghei/claude-skills --skill data-analyst
data-scientist	Full statistical pipeline: EDA, A/B testing, hypothesis testing, causal inference, ML model building, time series forecasting, SHAP/LIME interpretability.	skills.sh	npx skills add 404kidwiz/claude-supercode-skills --skill data-scientist

What makes these useful: The data-analyst skill from borghei encodes the "What/So What/Now What" reporting framework. Every analysis output includes the finding, the business impact, and the recommended action. That alone saves 30 minutes per report.

---

Stage 2: Data Cleaning & Transformation

Skill	What It Does	Source	Install
csv-data-wrangler	CSV processing expert: tool selection by file size (pandas vs polars vs DuckDB vs Spark), encoding/delimiter handling, data cleaning pipelines, SQL querying via DuckDB.	skills.sh	npx skills add 404kidwiz/claude-supercode-skills --skill csv-data-wrangler
data-cleaning-pipeline	Systematic cleaning: schema validation, missing value imputation (mean/median/predictive), duplicate removal, outlier detection (IQR/Z-score), type standardization, data quality scoring.	skills.sh	npx skills add aj-geddes/useful-ai-prompts --skill data-cleaning-pipeline
data-transform	Universal data transforms with pandas, numpy, sklearn: normalization, scaling (StandardScaler, MinMaxScaler, RobustScaler), encoding (LabelEncoder, OneHotEncoder), reshaping.	Microck/ordinary-claude-skills	Clone repo

What makes these useful: The csv-data-wrangler skill automatically selects the right tool for your file size. Under 1GB? pandas. 1-10GB? polars or DuckDB. Over 10GB? Spark. That decision tree alone prevents the #1 mistake data scientists make: loading a 5GB CSV into pandas and wondering why their laptop catches fire.

---

Stage 3: Visualization

Skill	What It Does	Source	Install
scientific-visualization	Publication-quality figures: matplotlib/seaborn/plotly, multi-panel layouts, error bars, significance markers, colorblind-safe palettes. Journal presets for Nature, Science, Cell. Export PDF/EPS/TIFF.	K-Dense-AI/scientific-agent-skills	Same repo
matplotlib (tvhahn)	9 opinionated chart patterns: horizontal bar, violin+strip, lollipop, decision boundary, heatmap, PR/ROC curves. DejaVu Sans, cubehelix/ColorBrewer palettes. Includes 10 source datasets + 29 public datasets.	tvhahn/matplotlib-skill	Clone repo
interactive-dashboard-builder	Self-contained HTML/JS dashboards with Chart.js, filters, interactivity, and professional styling. By Anthropic. 591 installs.	skills.sh	npx skills add anthropics/knowledge-work-plugins --skill interactive-dashboard-builder
visualization-expert	Chart selection guide, dashboard design, visual communication best practices. 2.9K installs.	skills.sh	npx skills add shubhamsaboo/awesome-llm-apps --skill visualization-expert
developing-with-streamlit	Official Streamlit skill: bundled reference docs for dashboards, themes, layouts, session state, custom components. Auto-discovers project environment. 1.8K installs.	streamlit/agent-skills	npx skills add streamlit/agent-skills --skill developing-with-streamlit

What makes these useful: The tvhahn matplotlib skill is remarkable. 9 opinionated chart patterns with journal-ready defaults. It includes 10 bundled datasets and references 29 public datasets. The Nature/Science/Cell presets for the scientific-visualization skill mean no more fighting with rcParams to get publication formatting right.

---

Stage 4: SQL & Query Generation

Skill	What It Does	Source	Install
sql-queries	Production SQL across 5 major dialects (PostgreSQL, Snowflake, BigQuery, Redshift, Databricks): window functions, CTEs, cohort retention, funnel analysis, EXPLAIN ANALYZE. By Anthropic.	skills.sh	npx skills add anthropics/knowledge-work-plugins --skill sql-queries
dbt integration	Translates business questions to dbt models: semantic layer first, then SQL modification, model discovery. Builds ref() and source() properly. Official dbt Labs skill.	dbt-labs/dbt-agent-skills	npx skills add dbt-labs/dbt-agent-skills
motherduck-query	DuckDB SQL for MotherDuck: write, validate, optimize analytical queries against cloud DuckDB. Part of 17-skill analytics catalog.	motherduckdb/agent-skills	Clone repo

What makes these useful: Anthropic's sql-queries skill covers 5 dialects with dialect-specific optimization hints. It knows that Snowflake prefers QUALIFY over subqueries, that BigQuery needs UNNEST for arrays, and that Redshift's DISTKEY matters for join performance. The dbt Labs skill is official - it understands ref() and source() natively.

---

Stage 5: Machine Learning & Modeling

Skill	What It Does	Source	Install
scikit-learn	Comprehensive classical ML: classification, regression, clustering, dimensionality reduction, preprocessing, hyperparameter tuning, production pipelines.	davila7/claude-code-templates	Clone repo
ai-ml-data-science	Full ML engineering: drift detection, leakage prevention, baselines-first model selection, slice analysis. LightGBM, CatBoost, sklearn, PyTorch, Polars.	skills.sh	npx skills add vasilyu1983/ai-agents-public --skill ai-ml-data-science
ml-pipeline-workflow	End-to-end MLOps: DAG orchestration (Airflow, Dagster, Kubeflow), experiment tracking, model versioning, deployment automation.	skills.sh	npx skills add wshobson/agents --skill ml-pipeline-workflow

The ai-ml-data-science skill enforces "baselines first" - it always starts with a simple model (logistic regression, random forest) before reaching for gradient boosting or neural nets. It also checks for data leakage automatically. Two mistakes that waste more data science time than any other.

---

Stage 6: Jupyter Notebooks

Skill	What It Does	Source	Install
jupyter-notebook	Clean, reproducible notebooks. Bundled templates + helper script to avoid JSON mistakes. Structured sections, metadata, reproducibility patterns. Official OpenAI skill. 2.4K installs.	openai/skills	npx skills add openai/skills --skill jupyter-notebook
working-in-notebooks	Jupyter, JupyterLab, and marimo notebook workflows.	legout/data-platform-agent-skills	Clone repo

The OpenAI jupyter-notebook skill is the one to start with. It includes bundled templates and a helper script that prevents the #1 notebook mistake: corrupting the .ipynb JSON structure by hand-editing cells.

---

Stage 7: Data Pipelines & ETL

Skill	What It Does	Source	Install
building-data-pipelines	Core batch ETL with Polars, DuckDB, PyArrow.	legout/data-platform-agent-skills	Clone repo
building-streaming-pipelines	Streaming ETL and real-time data processing.	legout/data-platform-agent-skills	Same repo
data-engineering-data-pipeline	ETL/ELT, Lambda, Kappa, Lakehouse architectures. Airflow/Prefect orchestration, dbt/Spark transforms, Delta Lake/Iceberg storage.	skills.sh	npx skills add rmyndharis/antigravity-skills
assuring-data-pipelines	Data quality and observability: Great Expectations, OpenTelemetry.	legout/data-platform-agent-skills	Same repo

The legout repo is a complete data platform in skill form: 14 skills covering ETL, streaming, quality, notebooks, ML, and app building. If you build data pipelines (not just analyze data), install this entire collection.

---

Stage 8: Document Output (xlsx, pdf, pptx, docx)

Source: anthropics/skills (Official, bundled with Claude)

Install:

/plugin install document-skills@anthropic-agent-skills

• xlsx: Programmatic Excel generation. Multi-sheet workbooks with pivot tables, charts, conditional formatting. Eliminates the "can you put that in a spreadsheet?" request.
• pdf: Extract text, tables, and structured data from research papers. Parse vendor reports and compliance documents.
• pptx: Generate the entire stakeholder presentation from your analysis. Charts, key metrics, methodology, recommendations.
• docx: Formal analysis reports with executive summary, methodology section, results tables, and appendices.

---

Stage 9: Scientific Research (144 skills)

K-Dense-AI/scientific-agent-skills is the largest scientific skill library at 27.9K stars. It covers:

• 70+ Python package skills: RDKit, Scanpy, BioPython, PyTorch Lightning, and more
• 78+ scientific databases: UniProt, PDB, ChEMBL, TCGA, and more
• Domain-specific workflows: RNA-seq pipelines, single-cell analysis, drug discovery, molecular dynamics, geospatial science, time series forecasting, quantum computing (PennyLane, Qiskit)

If you do computational biology, chemistry, or physics, this repo alone is worth the install.

---

Which Repos Have the Best Data Science Skills?

Repo	Stars	Skills	Focus
K-Dense-AI/scientific-agent-skills	27.9K	144	Biology, chemistry, physics, genomics, drug discovery
davila7/claude-code-templates	27.8K	20+ scientific	Scientific Python suite (matplotlib, sklearn, dask, scipy)
anthropics/skills	149K	21+	Document skills (xlsx, pdf, docx, pptx) + dashboard builder + SQL
legout/data-platform-agent-skills	-	14	Data engineering platform (ETL, ML, notebooks, quality)
motherduckdb/agent-skills	-	17	DuckDB/MotherDuck analytics workflows
tvhahn/matplotlib-skill	-	1	Publication-quality matplotlib (9 chart patterns, 39 datasets)
streamlit/agent-skills	-	1	Official Streamlit dashboard skill
dbt-labs/dbt-agent-skills	-	1+	Official dbt integration

---

What Is Missing? Building a Custom EDA Skill

Generic EDA skills exist (see Stage 1 above). But the biggest value comes from building one tuned to YOUR data. Here is the template:

---
name: eda
description: Run exploratory data analysis on tabular datasets. Use when the 
  user loads a CSV, Parquet, or database table and wants to understand its 
  structure, distributions, missing values, correlations, and anomalies. 
  Produces a formatted analysis report with visualizations.
dependencies: python>=3.10, pandas>=2.0, matplotlib, seaborn, scipy
---

What the skill should encode:

1. Schema discovery. dtypes, cardinality, sample values for every column. Classify columns as numeric, categorical, datetime, identifier, or free text.
2. Missing value analysis. Not just counts. Patterns of missingness (MCAR, MAR, MNAR). Recommendations for imputation strategy.
3. Distribution analysis. Histograms for numerics, value counts for categoricals, time series plots for datetime columns. Flag skew, outliers, and multimodal distributions.
4. Correlation analysis. Numeric correlations (Pearson and Spearman). Categorical associations (Cramér's V). Feature importance ranking.
5. Data quality checks. Duplicates, impossible values (negative ages, future dates), inconsistent categories ("US" vs "United States" vs "USA").
6. Gotchas list. Domain-specific corrections: "Revenue columns in this company are always in cents, not dollars." "The 'status' column uses legacy codes: 1=active, 2=churned, 3=paused."

Why build it yourself instead of waiting for the ecosystem: Your data dictionary is your competitive moat. A generic EDA skill runs the same analysis on every dataset. Your custom EDA skill knows that mrr means monthly recurring revenue, that plan_id=3 is the Enterprise tier, and that the created_at column has a known gap from the migration on 2025-03-15.

Use the skill-creator meta-skill to scaffold it:

"Create an EDA skill that analyzes tabular data. It should use pandas for 
data loading, seaborn for visualizations, and scipy for statistical tests. 
Include a gotchas list for common data quality issues."

Then customize the gotchas list with your domain knowledge. That is the skill nobody else can build for you.

---

How Do You Stack Skills for a Data Science Workflow?

Data ingestion:
  → composio (connect to databases, APIs, cloud storage)
  → sql-queries (Anthropic - 5 SQL dialects with optimization hints)
  → dbt integration (official dbt Labs - semantic layer, ref/source)
  → pdf skill (extract tables from research papers and reports)

Exploration & cleaning:
  → exploratory-data-analysis (K-Dense-AI - 200+ file formats)
  → data-analyst (borghei - What/So What/Now What reporting)
  → csv-data-wrangler (auto-selects pandas vs polars vs DuckDB by file size)
  → data-cleaning-pipeline (schema validation, imputation, dedup)

Analysis & modeling:
  → data-scientist (hypothesis testing, causal inference, SHAP/LIME)
  → scikit-learn (davila7 - full classical ML reference)
  → ai-ml-data-science (drift detection, leakage prevention, MLOps)

Visualization & dashboards:
  → scientific-visualization (publication-quality, journal presets)
  → matplotlib (tvhahn - 9 chart patterns, 39 datasets)
  → developing-with-streamlit (official Streamlit skill)
  → interactive-dashboard-builder (Anthropic - HTML/JS dashboards)

Reporting & output:
  → xlsx skill (multi-sheet workbooks, pivot tables)
  → pptx skill (stakeholder presentations from results)
  → docx skill (formal analysis reports)
  → jupyter-notebook (official OpenAI - clean reproducible notebooks)

---

How Does Jupyter Integration Work Today?

Claude Code operates in the terminal, not inside Jupyter. But the workflows compose well:

1. Claude Code generates the notebook. Ask it to create a .ipynb file with your analysis pipeline. It writes the cells, installs the dependencies, and structures the notebook with markdown explanations.
2. Claude Code generates scripts that notebooks import. Put your data transforms, feature engineering, and model evaluation code in .py files. Import them from notebooks. Claude Code maintains the Python files; you interact with the analysis through Jupyter.
3. Claude Code processes notebook output. Export results from Jupyter as CSVs or DataFrames saved to disk. Claude Code picks them up for report generation (xlsx, pptx, docx).

The gap is real-time notebook interaction. Claude Code cannot execute cells inside a running Jupyter kernel. But for the bookends (setup and reporting), skills eliminate the manual work.

---

Which MCP Servers Pair With Data Science Skills?

Skills tell the agent what to do. MCP servers give it access to data. The combination is where the real power is.

MCP Server	What It Connects To	Pairs With
Postgres MCP	PostgreSQL databases	EDA skill, scientific-agent-skills
BigQuery MCP	Google BigQuery	xlsx skill for report generation
Snowflake MCP	Snowflake data warehouse	Any analysis skill
S3 MCP	AWS S3 buckets	pdf skill for document processing
Composio	100+ apps (Stripe, Salesforce, HubSpot)	All data skills

---

Frequently Asked Questions

Can Claude Code replace Jupyter notebooks?

No. Claude Code and Jupyter serve different purposes. Jupyter is an interactive exploration environment with rich output rendering. Claude Code is a task-execution agent that generates, modifies, and processes files. They work best together: Claude Code generates notebooks and scripts, Jupyter executes and visualizes, Claude Code processes the results into reports.

What about large datasets?

Claude Code processes files on your local machine. It can handle datasets that fit in memory using pandas, or use chunked processing for larger files. For datasets that require distributed compute (Spark, Dask), use Claude Code to generate the pipeline code, then execute it in your cluster environment.

Do I need to know Python to use data science skills?

Not for the document skills (xlsx, pdf, pptx, docx) - those produce output files directly. For skills that generate Python code (scientific-agent-skills, custom EDA), basic Python literacy helps for reviewing and debugging the output. But the skill handles the code generation; you describe what you want in natural language.

Can skills connect to my company's databases?

Not directly. Skills are instruction files. Use an MCP server (Postgres MCP, BigQuery MCP, Snowflake MCP) to provide the database connection. The skill tells the agent what analysis to run; the MCP server provides the data access.

How do I share data science skills with my team?

Commit the skill folder to your repo's .claude/skills/ or .cursor/skills/ directory. Everyone who clones the repo gets the skill. For org-wide skills, use the Claude Plugins Marketplace or a shared GitHub repo. Your data dictionary and domain-specific gotchas should be in the skill - they are the most valuable part to share.

---

Sources

1. Anthropic Agent Skills Repository - Official document skills (xlsx, pdf, pptx, docx) and example skills
2. K-Dense-AI/scientific-agent-skills - Research, science, and finance workflow skills
3. composiohq/composio - 100+ app connectors with managed authentication
4. Agent Skills Specification - The open standard for portable AI agent capabilities
5. VoltAgent/awesome-agent-skills - Directory of 1,424+ curated skills

---

Related Content

• Anthropic Built 300+ Claude Code Skills Internally. Here's What They Learned. - How to build your own skills + the complete skills directory by category
• Claude Code for Sales - 60+ sales skills mapped to every pipeline stage
• Claude Code for Growth & Marketing - 120+ marketing skills with the shared context pattern
• Claude Code for DevOps - 150+ skills for Docker, K8s, Terraform, and CI/CD
• Claude Code for Solopreneurs - The 15-skill stack that replaces 5 hires

---

Start Here

Install the Document Skills first (/plugin install document-skills). Generate an Excel dashboard from your next analysis. Then build a custom EDA skill with your team's data dictionary and domain gotchas. That is the skill nobody else can build for you.

Want to build your own data science skills from scratch? The skills in this guide are starting points. The real value is encoding YOUR data dictionary, YOUR analysis patterns, YOUR domain-specific gotchas. Agent Skills 101 covers the full SKILL.md spec, progressive loading architecture, security best practices, and how to build skills that work across Claude Code, Cursor, and 50+ other agents. 25 lessons.

Start Agent Skills 101 →