tabular-eda-skill - SKILL.md Agent Skill

name: tabular-eda-skill description: Use for deterministic-first exploratory data analysis of tabular datasets with EDA Copilot, including configuring EDAConfig, running or inspecting run_eda outputs, reviewing evidence packets and artifacts, and drafting evidence-only explanations without raw rows or sensitive row-level values.

Tabular EDA Skill

Use this skill when the task involves repeatable exploratory data analysis for a tabular dataset, especially when the user needs data quality review, response analysis, feature relationship triage, leakage checks, drift checks, preprocessing recipe review, modeling-readiness review, exported artifacts, or evidence-grounded reporting.

Core Rules

Treat eda_copilot.core.workflow.run_eda as the deterministic source of truth.
Map user options to eda_copilot.core.config.EDAConfig; do not invent configuration fields.
Keep Streamlit as an interface only. Do not place deterministic calculations in UI code.
Any AI-facing explanation must use evidence-only context. Do not inspect or send raw rows, row samples, ID values, sensitive row-level values, or private data to AI interpretation.
If evidence is missing, say what is missing and use the fallback behavior rather than inventing metrics, columns, or conclusions.
Do not create a second EDA engine. Prefer wrappers and drilldowns over existing evidence packet sections before adding new calculations.
Treat preprocessing recipes as metadata. Processed row-level data must not be sent to AI interpretation.
Treat interaction exploration as exploratory two-feature segmentation only, not production modeling or model approval.

Workflow

Confirm the dataset is tabular and the requested analysis can be expressed through EDAConfig and existing EDA Copilot outputs.
Run or inspect the deterministic workflow described in actions/run_full_eda.md.
Review evidence packet sections, including analysis_objective, analyst_plan, eda_suites, recommended_drilldowns, eda_brief, and modeling_readiness_review, using references/current_contracts.md.
Use deterministic helper actions for EDA briefs, readiness, recommended drilldowns, bivariate drilldowns, preprocessing recipe metadata, interaction exploration, and report export when available.
For explanation or drafting tasks, use actions/review_evidence_packet.md and keep every claim tied to packet fields.
Apply fallback.md when input data, configuration, evidence, or optional AI support is unavailable.
Before handoff, use validation_checklist.md.

Bundled References

actions/run_full_eda.md: action contract for deterministic full EDA runs.
actions/build_eda_brief.md: action contract for deterministic EDA Briefs.
actions/build_modeling_readiness_review.md: action contract for readiness review.
actions/build_feature_drilldown.md: action contract for feature drilldowns.
actions/build_bivariate_drilldown.md: action contract for pair drilldowns.
actions/apply_preprocessing_recipe.md: action contract for preprocessing recipes.
actions/run_interaction_explorer.md: action contract for exploratory two-feature segmentation.
actions/export_eda_report.md: action contract for report export.
actions/review_evidence_packet.md: action contract for evidence-only review and explanation.
schemas/*.schema.json: static documentation-level contracts for action inputs, outputs, and required evidence packet keys.
examples/basic_binary_eda.md: minimal example request and expected handling.
references/current_contracts.md: current EDAConfig, evidence packet, and artifact baseline.
fallback.md: required fallback behavior.
validation_checklist.md: completion checklist.