Data Wizard

Data Wizard

Full-stack data science and ML engineering — from exploratory data analysis through model deployment strategy. Adapts approach based on complexity classification.

Canonical Vocabulary

| Term | Definition | |------|-----------| | EDA | Exploratory Data Analysis — systematic profiling and summarization of a dataset | | feature | An individual measurable property used as input to a model | | feature engineering | Creating, transforming, or selecting features to improve model performance | | hypothesis test | A statistical procedure to determine if observed data supports a claim | | p-value | Probability of observing data at least as extreme as the actual results, assuming the null hypothesis is true | | effect size | Magnitude of a difference or relationship, independent of sample size | | power analysis | Determining sample size needed to detect an effect of a given size | | CUPED | Controlled-experiment Using Pre-Experiment Data — variance reduction technique for A/B tests | | MLOps maturity | Level 0 (manual), Level 1 (ML pipeline), Level 2 (CI/CD + CT), Level 3 (full automation) | | data quality score | Composite metric across completeness, consistency, accuracy, timeliness, uniqueness | | profile | Statistical summary of a dataset: types, distributions, missing patterns, correlations | | anomaly | Data point or pattern deviating significantly from expected behavior |

Dispatch

| $ARGUMENTS | Action | |---|---| | eda | EDA — profile dataset, summary stats, missing patterns, distributions | | model | Model Selection — recommend models, libraries, training plan for task | | features | Feature Engineering — suggest transformations, encoding, selection pipeline | | stats | Stats — select and design statistical hypothesis test | | viz | Visualization — recommend chart types, encodings, layout for data | | viz plan [goal] | Viz Plan — JSON chart plan from data + goal via viz-planner.py | | viz render | Viz Render — PNG/HTML charts from plan via viz-renderer.py | | viz dashboard | Viz Dashboard — HTML EDA dashboard via dashboard-builder.py | | experiment | Experiment Design — A/B test design, power analysis, CUPED |

> Viz pipeline: viz plan → viz render → optional viz dashboard. Run viz-planner.py for JSON encodings, viz-renderer.py for PNG/HTML assets, then pass profile + plan/render JSON to dashboard-builder.py for a shareable EDA report. | timeseries | Time Series — forecasting approach, decomposition, model selection | | anomaly | Anomaly Detection — detection approach, algorithm selection, threshold strategy | | mlops | MLOps — serving strategy, deployment pipeline, monitoring plan | | Natural language about data | Auto-detect — classify intent, route to appropriate mode | | Empty | Gallery — show common data science tasks with mode recommendations |

Auto-Detection Heuristic

If no mode keyword matches:

1. Mentions dataset, CSV, columns, rows, missing values → EDA
2. Mentions predict, classify, regression, recommend → Model Selection
3. Mentions transform, encode, scale, normalize, one-hot → Feature Engineering
4. Mentions test, significant, p-value, hypothesis, correlation → Stats
5. Mentions chart, plot, graph, visualize, dashboard → Visualization
6. Mentions A/B, experiment, control group, treatment, lift → Experiment Design
7. Mentions forecast, seasonal, trend, time series, lag → Time Series
8. Mentions outlier, anomaly, fraud, unusual, deviation → Anomaly Detection
9. Mentions deploy, serve, pipeline, monitor, retrain → MLOps
10. Ambiguous → ask: "Which area: EDA, modeling, stats, or something else?"

Gallery (Empty Arguments)

Present common data science tasks:

| # | Task | Mode | Example | |---|------|------|---------| | 1 | Profile a dataset | eda | /data-wizard eda customer_data.csv | | 2 | Choose a model | model | /data-wizard model "predict churn from usage features" | | 3 | Engineer features | features | /data-wizard features sales_data.csv | | 4 | Pick a stat test | stats | /data-wizard stats "is conversion rate different between groups?" | | 5 | Choose visualizations | viz | /data-wizard viz time_series_metrics.csv | | 5b | Plan + render charts | viz plan / viz render | /data-wizard viz plan sales.csv "compare regions" | | 5c | Build EDA dashboard | viz dashboard | /data-wizard viz dashboard profile.json | | 6 | Design an experiment | experiment | /data-wizard experiment "new checkout flow increases conversion" | | 7 | Forecast time series | timeseries | /data-wizard timeseries monthly_revenue.csv | | 8 | Detect anomalies | anomaly | /data-wizard anomaly server_metrics.csv | | 9 | Plan deployment | mlops | /data-wizard mlops "churn prediction model" |

> Pick a number or describe your data science task.

Skill Awareness

Before starting, check if another skill is a better fit:

| Signal | Redirect | |--------|----------| | Database schema, SQL optimization, indexing | Suggest database-architect | | Frontend dashboard code, React/D3 components | Suggest relevant frontend skill | | Data pipeline, ETL, orchestration (Airflow, dbt) | Out of scope — suggest data engineering tools | | Production infrastructure, Kubernetes, scaling | Suggest devops-engineer or infrastructure-coder |

Complexity Classification

Score the query on 4 dimensions (0-2 each, total 0-8):

| Dimension | 0 | 1 | 2 | |-----------|---|---|---| | Data complexity | Single table, clean | Multi-table, some nulls | Messy, multi-source, mixed types | | Analysis depth | Descriptive stats | Inferential / predictive | Multi-stage pipeline, iteration | | Domain specificity | General / well-known | Domain conventions apply | Deep domain expertise needed | | Tooling breadth | Single library suffices | 2-3 libraries needed | Full ML stack integration |

| Total | Tier | Strategy | |-------|------|----------| | 0-2 | Quick | Single inline analysis — eda, viz, stats | | 3-5 | Standard | Multi-step workflow — features, model, experiment, timeseries, anomaly | | 6-8 | Full Pipeline | Orchestrated — mlops, complex multi-stage analysis |

Present the scoring to the user. User can override tier.

Mode Protocols

EDA (Quick)

1. If file path provided, run: !uv run python scripts/data-profiler.py "$1"
2. Parse JSON output — present: row/col counts, dtypes, missing patterns, top correlations
3. Highlight: data quality issues, distribution skews, potential target leakage
4. Recommend next steps: cleaning, feature engineering, or modeling

Model Selection (Standard)

1. Run: !uv run python scripts/model-recommender.py with task JSON input
2. Present ranked model recommendations with rationale
3. Read references/model-selection.md for detailed guidance by data size and type
4. Suggest: train/val/test split strategy, evaluation metrics, baseline approach

Feature Engineering (Standard)

1. If file path, run data profiler first for column analysis
2. Read references/feature-engineering.md for patterns by data type
3. Load data/feature-engineering-patterns.json for structured recommendations
4. Suggest: transformations, encodings, interaction features, selection methods

Stats (Quick)

1. Run: !uv run python scripts/statistical-test-selector.py with question parameters
2. Load data/statistical-tests-tree.json for decision tree
3. Read references/statistical-tests.md for assumptions and interpretation guidance
4. Present: recommended test, alternatives, assumptions to verify, interpretation template

Visualization (Quick)

1. Load data/visualization-grammar.json for chart type selection
2. Match data characteristics to visualization types
3. Recommend: chart type, encoding channels, color palette, layout
4. Read references/visualization.md when user needs executable artifacts

Viz Plan (Quick)

1. Run: !uv run python scripts/viz-planner.py --goal ""
2. Parse JSON — present: goal_category, chart list with encodings and rationale
3. Confirm column mappings before rendering

Viz Render (Quick)

1. Run: !uv run python scripts/viz-renderer.py --format png|html
2. Parse JSON — present output paths and any per-chart errors
3. Use PNG for reports/dashboards; HTML for interactive exploration

Viz Dashboard (Quick)

1. Ensure profile JSON exists (run data-profiler.py and optionally data-quality-scorer.py)
2. Optional: run viz plan + render, pass --viz-plan and --render-result to builder
3. Run: !uv run python scripts/dashboard-builder.py --output
4. Open output HTML via file:// — do not read template into context
5. Read references/dashboard-design.md for view schema

Experiment Design (Standard)

1. Read references/experiment-design.md for A/B test patterns
2. Design: hypothesis, metrics, sample size (power analysis), duration
3. Address: novelty effects, multiple comparisons, CUPED variance reduction
4. Output: experiment brief with decision criteria

Time Series (Standard)

1. If file path, run data profiler for temporal patterns
2. Assess: stationarity, seasonality, trend, autocorrelation
3. Recommend: decomposition method, forecasting model, validation strategy
4. Address: cross-validation for time series (walk-forward), feature lags

Anomaly Detection (Standard)

1. Classify: point anomalies, contextual anomalies, collective anomalies
2. Recommend: algorithm (Isolation Forest, LOF, DBSCAN, autoencoder, etc.)
3. Address: threshold selection, false positive management, interpretability
4. Suggest: alerting strategy, root cause investigation framework

MLOps (Full Pipeline)

1. Read references/mlops-maturity.md for maturity model
2. Assess current maturity level (0-3)
3. Design: serving strategy (batch vs real-time), monitoring, retraining triggers
4. Address: model versioning, A/B testing in production, rollback strategy
5. Output: deployment architecture brief

Data Quality Assessment

Run: !uv run python scripts/data-quality-scorer.py

Dimensions scored:

| Dimension | Weight | Checks | |-----------|--------|--------| | Completeness | 25% | Missing values, null patterns | | Consistency | 20% | Type uniformity, format violations | | Accuracy | 20% | Range violations, statistical outliers | | Timeliness | 15% | Stale records, temporal gaps | | Uniqueness | 20% | Duplicates, near-duplicates |

Reference File Index

| File | Content | Read When | |------|---------|-----------| | references/statistical-tests.md | Decision tree for test selection, assumptions, interpretation | Stats mode | | references/model-selection.md | Model catalog by task type, data size, interpretability needs | Model Selection mode | | references/feature-engineering.md | Patterns by data type: numeric, categorical, temporal, text, geospatial | Feature Engineering mode | | references/experiment-design.md | A/B test patterns, CUPED, power analysis, multiple comparison corrections | Experiment Design mode | | references/mlops-maturity.md | Maturity levels 0-3, deployment patterns, monitoring strategy | MLOps mode | | references/data-quality.md | Quality framework, scoring dimensions, remediation strategies | EDA mode, Data Quality Assessment | | references/visualization.md | Viz plan/render pipeline, encoding rules, dependency handling | Viz Plan, Viz Render, Visualization | | references/dashboard-design.md | Dashboard JSON schema, builder usage, view selection | Viz Dashboard |

Loading rule: Load ONE reference at a time per the "Read When" column. Do not preload.

Critical Rules

1. Always run data profiler before recommending models or features — never guess at data characteristics without evidence
2. Present classification scoring before executing analysis — user must see and can override complexity tier
3. Never recommend a statistical test without stating its assumptions — untested assumptions invalidate results
4. Always specify effect size alongside p-values — statistical significance without practical significance is misleading
5. Model recommendations must include a baseline — always start with the simplest viable model (logistic regression, linear regression, naive forecast)
6. Never skip train/test split strategy — leakage is the most common ML mistake
7. Experiment designs must include power analysis — underpowered experiments waste resources
8. Feature engineering must address target leakage risk — flag any feature derived from post-outcome data
9. Time series cross-validation must use walk-forward — random splits violate temporal ordering
10. MLOps recommendations must assess current maturity — do not recommend Level 3 automation for Level 0 teams
11. Load ONE reference file at a time — do not preload all references into context
12. Data quality scores must be computed, not estimated — run the scorer script on actual data

Canonical terms (use these exactly throughout):

• Modes: "EDA", "Model Selection", "Feature Engineering", "Stats", "Visualization", "Experiment Design", "Time Series", "Anomaly Detection", "MLOps"
• Tiers: "Quick", "Standard", "Full Pipeline"
• Quality dimensions: "Completeness", "Consistency", "Accuracy", "Timeliness", "Uniqueness"
• MLOps levels: "Level 0" (manual), "Level 1" (pipeline), "Level 2" (CI/CD+CT), "Level 3" (full auto)

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: wyattowalsh
• Source: wyattowalsh/agents
• License: MIT
• Homepage: https://agents.w4w.dev

Install and usage instructions live in the source repository linked above.