Data Science Agent Skills

Twenty-five independent, production-grade agent skills for data science work where judgment matters: framing, readiness, data quality, leakage, reproducibility, evaluation, governance, privacy, MLOps, incident response, and stakeholder communication.

This repository is designed as a public skill catalog, not a single monolithic toolkit. Every skill lives on its own page, has its own assets, and can be copied or installed independently.

What This Repository Gives You

• Independent skill pages: each folder has a public README.md, portable SKILL.md, manifest, references, scripts, tests, hero shot, and detailed infographic.
• Agent-ready workflows: each skill tells the agent when to use it, what evidence to request, what checks to run, and what artifact to return.
• Provider guidance: the skills are written to work across Codex, ChatGPT Agents, Claude, Gemini, Copilot-style agents, Cursor, Windsurf, Gravity, LangGraph, CrewAI, AutoGen, OpenHands, Goose, OpenCode, and local LLM agents.
• Decision-grade outputs: every skill returns the same core contract: Problem, Inputs, Checks performed, Findings, Decision, Risks, Next actions, and Artifacts.

Pick A Skill By Pain Point

Pain point	Start here	What the skill helps produce
The business question is vague or not measurable.	Data Question Framing	Decision canvas, target definition, metric tree, constraints, decision owner
A team wants AI before the data is ready.	AI Data Readiness Triage	Readiness scorecard, blocker list, owner map, go/no-go recommendation
Schemas are messy and no practical data contract exists.	Dataset Contract Forensics	Contract draft, quality checks, Pandera/Great Expectations/dbt test suggestions
Nobody can explain where fields, metrics, or features came from.	Semantic Lineage Mapper	Lineage map, dependency risk list, owner prompts, evidence ledger
Cleaning is happening silently and cannot be audited.	Messy Data Cleanroom	Reversible cleaning plan, mutation log, reviewable assumptions
A model may be leaking target, time, group, split, or join information.	Leakage Adversary	Leakage attack report, split validation notes, contamination fixes
Prediction work is being sold as causal proof.	Causal Assumption Scout	DAG, confounder checklist, negative controls, sensitivity plan
The dataset is too small for confident modeling.	Small Data Power Planner	Power plan, uncertainty statement, resampling/Bayesian guidance, stop rule
Missing values are being imputed without mechanism review.	Missingness Mechanism Investigator	MCAR/MAR/MNAR assessment, missingness profile, sensitivity plan
A notebook only works on one machine or one execution order.	Notebook Repro Packager	Environment capture, data manifest, seeds, smoke test, repro checklist
Exploratory notebooks need to become scripts or DAG tasks.	Notebook To Pipeline Surgeon	Refactor plan, parameter map, CLI contract, test seams, pipeline options
Experiments cannot be audited or compared later.	Experiment Provenance Ledger	Code/data/env/model/metric ledger, artifact checklist, decision trace
Complex models are being trusted without baselines or ablations.	Baseline Ablation Lab	Baseline matrix, ablation plan, lift evidence, complexity verdict
Metrics are disconnected from real decisions and costs.	Metric Design Review Board	Metric tree, guardrails, threshold/cost curves, review verdict
Evaluation does not match deployment reality.	Evaluation Reality Check	Split strategy, backtest plan, external validity checks, uncertainty report
Governance artifacts are missing before release.	Model Risk Compliance Memo	Privacy, fairness, safety, auditability, monitoring, and approval memo
Sensitive data is being handled with too much exposure.	Privacy Preserving Data Science	PII discovery, minimization plan, privacy technique decision tree
Dependencies, notebooks, model weights, or datasets may create supply-chain risk.	OSS Data Science Supply Chain Auditor	SBOM-style risk report, license/provenance notes, remediation plan
Cloud or GPU spend is running ahead of evidence.	Cost Aware Compute Planner	Local-vs-cloud plan, sampling ladder, budget guardrails, stop rules
A prototype needs to be handed to engineering.	Production ML Handoff	API contract, artifact registry notes, rollback plan, SLOs, monitoring owners
Production data or model behavior may drift.	Drift Monitor Designer	Drift metrics, thresholds, label-delay handling, alert/retrain playbook
A feature store may be useful, but could add complexity or skew.	Feature Store Readiness	Store-or-skip decision, feature contract, TTL, lineage, skew checks
A model, dashboard, or data product broke and needs RCA.	Data Science Incident RCA	Timeline, blast radius, root cause, countermeasures, prevention plan
Analysis needs to become a decision-grade story.	Stakeholder Insight Storyteller	Executive memo, visual hierarchy, uncertainty language, recommendation
Domain experts need to challenge agent assumptions.	Domain Expert Collaboration Loop	SME interview prompts, assumption log, evidence ledger, signoff loop

Skill Map

The portfolio is intentionally focused on gaps that generic data-analysis tools do not solve well by themselves: judgment, provenance, readiness, leakage, domain review, governance, and operational handoff.

How To Use A Skill

1. Open the skill folder in data-science-agent-skills/.
2. Read that skill's independent README.md to understand the problem, inputs, workflow, outputs, and validation.
3. Copy the whole skill folder into your agent's skill directory or reference the folder from your agent runtime.
4. Give the agent the relevant data, notebooks, SQL, model artifacts, stakeholder context, or incident evidence.
5. Keep the returned artifact with the project so future humans and agents can audit the decision.

Agent And Provider Compatibility

The skills avoid provider-specific magic. They are written as portable workflows with explicit inputs, checks, final artifacts, and failure conditions.

Provider or agent style	How to use these skills
OpenAI Codex and ChatGPT Agents	Copy the skill folder into the skills directory, then invoke the skill by name or by matching the task trigger.
Claude and Claude Code	Paste or install the SKILL.md workflow and keep the references/, scripts/, and tests/ folders available as local context.
Gemini, Copilot, Cursor, Windsurf, Gravity, Goose, OpenCode, OpenHands	Treat each folder as a portable playbook: read README.md, follow SKILL.md, run deterministic scripts when available, and return the standard artifact.
LangGraph, CrewAI, AutoGen, LlamaIndex, Semantic Kernel	Wrap the skill as a node/tool with the declared inputs and output contract from MANIFEST.json.
Local LLM agents	Keep raw evidence local, run the included tests/scripts locally, and use the README workflow as the orchestration policy.

Provider output is evidence, not authority. The agent should inspect local artifacts first, preserve raw evidence, avoid silent mutation, state uncertainty, and ask for user or owner confirmation when missing context would change the decision.

What Every Independent Skill Page Includes

File or folder	Purpose
SKILL.md	Portable agent instructions, trigger guidance, workflow, checks, and output contract
README.md	Public page explaining the problem, when to use the skill, inputs, workflow, examples, provider notes, and validation
MANIFEST.json	Machine-readable package metadata and contracts
references/	Playbooks, interoperability notes, research grounding, rubrics, and test guidance
scripts/	Deterministic helper scripts where the workflow benefits from repeatable checks
tests/	Contract tests plus clean, messy, and adversarial fixture coverage
assets/hero-shot.png	Skill-specific 1600x900 hero image
assets/reddit-infographic.png	Skill-specific 1080x1350 infographic with detailed workflow explanation

Full Independent Skill Index

Skill	What it solves	Main artifact
AI Data Readiness Triage	Scores whether data is ready for AI or analytics before modeling starts.	Readiness scorecard and blocker plan
Baseline Ablation Lab	Forces honest baselines and ablations before trusting complex models.	Baseline table and ablation matrix
Causal Assumption Scout	Stops prediction work from being oversold as causal evidence.	DAG, assumption map, sensitivity notes
Cost Aware Compute Planner	Prevents runaway cloud and GPU spend before experiments launch.	Compute budget plan and stop rules
Data Question Framing	Turns vague stakeholder asks into measurable data-science problems.	Decision canvas and metric tree
Data Science Incident RCA	Diagnoses broken models, dashboards, data products, and pipeline incidents.	RCA memo, timeline, countermeasures
Dataset Contract Forensics	Infers practical data contracts and tests from messy schemas and samples.	Contract draft and test suggestions
Domain Expert Collaboration Loop	Structures SME review so agents do not miss domain assumptions.	Assumption log and signoff trail
Drift Monitor Designer	Designs data, concept, and performance drift monitoring with actions.	Monitoring design and alert playbook
Evaluation Reality Check	Builds trustworthy evaluation for tabular, time-series, NLP, CV, and ranking.	Evaluation plan and stress-test report
Experiment Provenance Ledger	Preserves code, data, environment, model, metrics, and decisions.	Experiment ledger and audit trail
Feature Store Readiness	Decides whether a feature store is warranted and how to design features safely.	Store-or-skip decision and feature contract
Leakage Adversary	Attacks ML pipelines for target, temporal, group, split, and join leakage.	Leakage report and split validation notes
Messy Data Cleanroom	Produces reversible cleaning plans instead of silent data edits.	Cleaning plan and audit log
Metric Design Review Board	Aligns metrics with real decisions, costs, thresholds, and guardrails.	Metric tree and review verdict
Missingness Mechanism Investigator	Diagnoses MCAR, MAR, MNAR, or unknown missingness before imputation.	Missingness report and sensitivity plan
Model Risk Compliance Memo	Creates governance-ready artifacts for privacy, fairness, safety, auditability, and monitoring.	Model risk memo and conditions
Notebook Repro Packager	Makes notebooks runnable elsewhere with environment, seeds, data manifest, and smoke tests.	Repro package and smoke test
Notebook To Pipeline Surgeon	Refactors exploratory notebooks into modular scripts or DAG tasks.	Pipeline extraction plan and CLI contract
OSS Data Science Supply Chain Auditor	Audits dependencies, notebooks, model weights, datasets, licenses, and provenance.	SBOM-style risk report
Privacy Preserving Data Science	Guides minimized, local-first, aggregated, synthetic, federated, or enclave analysis.	Privacy decision tree and safe-analysis plan
Production ML Handoff	Turns data-science prototypes into engineering-ready production handoffs.	Handoff packet, API contract, monitoring plan
Semantic Lineage Mapper	Reconstructs lineage from SQL, notebooks, pipelines, configs, and dashboards.	Lineage map and owner prompts
Small Data Power Planner	Helps teams avoid overfitting, false certainty, and premature modeling with scarce data.	Power plan, uncertainty summary, stop rule
Stakeholder Insight Storyteller	Converts analysis into decision-grade narrative without overclaiming.	Executive memo and recommendation narrative

Validation

Run the contract checks inside any skill folder:

python scripts/quick_validate_skill.py . --strict
python tests/test_skill_contract.py .

For the full repository, repeat those checks across all folders in data-science-agent-skills/.