emily2040/data-science-agent-skills
Data science & MLIndependent Agent/Codex skills for high-leverage data science workflows, with hero shots, infographics, validation scripts, and provider interoperability notes.
Overview
A production-grade catalog of with deterministic Python standard-library verification tools, strict eight-section decision contracts, and native execution support. Every skill requires the same final decision structure (Problem, Inputs, Checks performed, Findings, Decision, Risks, Next actions, Artifacts) and emits an explicit five-state verdict (go, fix-first, stop, proceed-with-risk, needs-owner-review). All helper scripts and validators use only the Python 3 standard library—no pip install or virtual environment required. - : OpenAI Codex, ChatGPT Agents, Anthropic Claude, Claude Code, Google Gemini, Gravity / Antigravity, GitHub Copilot, Cursor, Windsurf, Goose, OpenCode, OpenHands, LangGraph, CrewAI, AutoGen, LlamaIndex, Semantic Kernel, and Local LLM Agents. - : Every skill includes references/multilingual-guide.md with localized section headers, decision verdicts, and domain terminology in , , , and .
README
Data Science Agent Skills
Language / 语言 / 言語 / 언어: English | 简体中文 | 日本語 | 한국어
A production-grade catalog of 25 judgment-heavy data science and ML engineering agent skills with deterministic Python standard-library verification tools, strict eight-section decision contracts, and native English, Simplified Chinese (简体中文), Japanese (日本語), and Korean (한국어) execution support.
Every skill requires the same final decision structure (Problem, Inputs, Checks performed, Findings, Decision, Risks, Next actions, Artifacts) and emits an explicit five-state verdict (go, fix-first, stop, proceed-with-risk, needs-owner-review).
Find The Right Skill By Project Stage
| Lifecycle Stage | Skill Package | Core Judgment & Scope | Primary Deterministic Script |
|---|---|---|---|
| Stage 1: Framing & Triage | Data Question Framing (data-question-framing) |
Turn vague business or research asks into measurable DS problem statements | scripts/decision_canvas_validator.py |
| Stage 1: Framing & Triage | AI Data Readiness Triage (ai-data-readiness-triage) |
Score whether datasets are ready for AI, analytics, or modeling | scripts/profile_table.py |
| Stage 1: Framing & Triage | Dataset Contract Forensics (dataset-contract-forensics) |
Infer and verify schema, grain, uniqueness, freshness, and null contracts | scripts/profile_table.py |
| Stage 1: Framing & Triage | Semantic Lineage Mapper (semantic-lineage-mapper) |
Trace table, column, and metric definitions across SQL, notebooks, and pipelines | scripts/extract_lineage_tokens.py |
| Stage 1: Framing & Triage | Messy Data Cleanroom (messy-data-cleanroom) |
Design reversible, auditable cleaning and quarantine rules without mutating raw data | scripts/profile_table.py |
| Stage 2: Validity & Leakage | Leakage Adversary (leakage-adversary) |
Red-team ML datasets and pipelines for target, temporal, group, and preprocessing leakage | scripts/split_leakage_probe.py |
| Stage 2: Validity & Leakage | Causal Assumption Scout (causal-assumption-scout) |
Audit causal claims for confounding, collider bias, post-treatment controls, and identification | scripts/confounder_backdoor_check.py |
| Stage 2: Validity & Leakage | Small Data Power Planner (small-data-power-planner) |
Calculate minimum detectable effect (MDE), sample power, and low-N overfitting bounds | scripts/power_sample_size.py |
| Stage 2: Validity & Leakage | Missingness Mechanism Investigator (missingness-mechanism-investigator) |
Diagnose MCAR, MAR, and MNAR missingness and block naive imputation traps | scripts/profile_table.py |
| Stage 2: Validity & Leakage | Baseline Ablation Lab (baseline-ablation-lab) |
Require naive/simple baselines and controlled ablations before accepting complex models | scripts/ablation_delta_audit.py |
| Stage 3: Repro & Pipelines | Notebook Repro Packager (notebook-repro-packager) |
Audit Jupyter notebooks for out-of-order cells, unseeded RNG, hardcoded paths, and lockfiles | scripts/notebook_repro_audit.py |
| Stage 3: Repro & Pipelines | Notebook To Pipeline Surgeon (notebook-to-pipeline-surgeon) |
Refactor exploratory notebooks into modular, testable, idempotent pipeline stages | scripts/notebook_repro_audit.py |
| Stage 3: Repro & Pipelines | Experiment Provenance Ledger (experiment-provenance-ledger) |
Record SHA-256 dataset hashes, git commit state, seeds, and hyperparameters per run | scripts/capture_provenance.py |
| Stage 3: Repro & Pipelines | OSS DS Supply Chain Auditor (oss-ds-supply-chain-auditor) |
Audit Python/DS dependency manifests for unpinned packages, URL installs, and pickle risks | scripts/dependency_surface.py |
| Stage 3: Repro & Pipelines | Cost Aware Compute Planner (cost-aware-compute-planner) |
Estimate retry-adjusted GPU/cloud spend and enforce 10% pilot gates before full runs | scripts/compute_budget_estimator.py |
| Stage 4: Eval & Governance | Metric Design Review Board (metric-design-review-board) |
Stress-test primary metrics, denominators, time windows, and anti-gaming guardrails | scripts/metric_guardrail_check.py |
| Stage 4: Eval & Governance | Evaluation Reality Check (evaluation-reality-check) |
Audit subgroup slice errors, probability calibration, and offline-to-business translation | scripts/slice_error_auditor.py |
| Stage 4: Eval & Governance | Model Risk Compliance Memo (model-risk-compliance-memo) |
Document blast radius, human-in-the-loop override, fallback policy, and governance gates | scripts/risk_tier_gate.py |
| Stage 4: Eval & Governance | Privacy Preserving DS (privacy-preserving-ds) |
Audit direct PII, quasi-identifiers, k-anonymity risks, and retention boundaries | scripts/profile_table.py |
| Stage 4: Eval & Governance | Domain Expert Collaboration Loop (domain-expert-collaboration-loop) |
Measure inter-annotator agreement (Cohen’s kappa) and resolve SME label taxonomy disputes | scripts/sme_label_agreement.py |
| Stage 5: Ops & Storytelling | Production ML Handoff (production-ml-handoff) |
Verify model SHA-256, input/output contracts, p99 latency SLA, fallback, and rollback readiness | scripts/handoff_readiness_audit.py |
| Stage 5: Ops & Storytelling | Feature Store Readiness (feature-store-readiness) |
Audit online-offline feature parity, point-in-time joins, and serving TTL freshness | scripts/online_offline_skew_check.py |
| Stage 5: Ops & Storytelling | Drift Monitor Designer (drift-monitor-designer) |
Detect mean shifts, variance explosions, constant-column breaks, NaNs, and missing-rate spikes | scripts/drift_probe.py |
| Stage 5: Ops & Storytelling | Data Science Incident RCA (data-science-incident-rca) |
Correlate broken dashboards or model drops across ingestion, schema, feature, model, and metric layers | scripts/incident_timeline_correlator.py |
| Stage 5: Ops & Storytelling | Stakeholder Insight Storyteller (stakeholder-insight-storyteller) |
Turn validated technical evidence into an executive decision memo with uncertainty and caveats | scripts/memo_claim_evidence_check.py |
Quickstart: Validate, List in 4 Languages, or Install Flat
All helper scripts and validators use only the Python 3 standard library—no pip install or virtual environment required.
# Validate all 25 skills (runs strict contract + behavioral script smoke tests)
python scripts/catalog_cli.py validate
# List the catalog in English, Simplified Chinese, Japanese, or Korean
python scripts/catalog_cli.py list --lang en
python scripts/catalog_cli.py list --lang zh-CN
python scripts/catalog_cli.py list --lang ja
python scripts/catalog_cli.py list --lang ko
# Install flat //SKILL.md bundles for single-level agent skill discovery
python scripts/catalog_cli.py install --target ~/.codex/skills
python scripts/catalog_cli.py install --target ~/.claude/skills
python scripts/catalog_cli.py install --target ~/.gemini/config/skills
# Run the end-to-end multi-skill lifecycle and regression test suite
python -m unittest discover -s tests -p "test_*.py" -v
Cross-Runtime & Cloud Provider Interoperability
- 18 Supported Agent Runtimes: OpenAI Codex, ChatGPT Agents, Anthropic Claude, Claude Code, Google Gemini, Gravity / Antigravity, GitHub Copilot, Cursor, Windsurf, Goose, OpenCode, OpenHands, LangGraph, CrewAI, AutoGen, LlamaIndex, Semantic Kernel, and Local LLM Agents.
- Native Quadrilingual Execution: Every skill includes references/multilingual-guide.md with localized section headers, decision verdicts, and domain terminology in English, 简体中文, 日本語, and 한국어.
- Executable Helper Scripts and Strict Contract Tests: Each skill bundles a domain-specific stdlib CLI verifier in
scripts/,scripts/checklist_score.py,scripts/quick_validate_skill.py, andtests/test_skill_contract.py.
Recommended Tools
Try a different keyword or remove a filter.
Install
npx skillfish add emily2040/data-science-agent-skills