Using data engineering agent skills

Using Data Engineering Agent Skills

Overview

Start here before changing code, SQL, orchestration, contracts, or infrastructure. This skill maps the user request to the right data engineering workflow so the agent does not skip specification, quality, governance, replay, or rollback thinking.

When to Use

• starting a new session
• deciding which skill should lead execution
• translating a vague request into the right workflow
• choosing the best preset, starter pack, or runnable example
• deciding the safest next command

Do not stop here once the task has been classified. Load the actual execution skill after triage.

Workflow

1. Classify the task type.

   • New data product, new pipeline, or major behavior change: use data-specification
   • Approved scope that needs sequencing: use pipeline-planning-and-task-breakdown
   • Implementation across code, SQL, or jobs: use the relevant build skill plus a matching preset
   • File drops, SFTP, or partner-managed feeds: use file-and-partner-feed-ingestion
   • Glue Data Catalog, Lake Formation, or AWS-native catalog governance: use glue-data-catalog-and-lake-formation-governance
   • Python implementation work: use python-data-engineering-and-pipeline-packaging
   • Scala JVM data jobs: use scala-data-engineering-on-jvm-runtimes
   • Java connectors or data services: use java-data-engineering-and-integration-services
   • Unity Catalog or Databricks lakehouse governance: use unity-catalog-and-lakehouse-governance
   • Purview or Azure-native governance: use microsoft-purview-and-azure-data-governance
   • Dataplex, policy tags, or BigQuery governance: use dataplex-and-bigquery-governance
   • MySQL versus NoSQL or operational-store choice: use operational-datastore-selection-relational-and-nosql
   • ETL, ELT, or transformation-boundary redesign: use etl-elt-and-modernization-strategy
   • Test data generation, seeded fixtures, or lower-environment realism: use test-data-preparation-and-synthetic-data
   • Production-like data refresh into development, QA, or staging: use lower-environment-data-masking-and-obfuscation
   • Quality, reconciliation, or contract correctness: use data-quality-and-contract-testing
   • Resiliency testing, failure injection, restart drills, or failover validation: use data-resiliency-testing-and-failure-injection
   • Disaster recovery or business continuity planning: use data-platform-disaster-recovery-and-business-continuity
   • Reliability issue, incident, or broken publish: use incident-triage-and-pipeline-recovery
   • Replay, rerun, cutover, or historical repair: use safe-backfill-and-replay-orchestration, orchestration-and-backfills, and data-migration-and-platform-cutover
   • Serverless Spark on Lambda or short-lived runtimes: use spark-serverless-reliability-and-state-management
   • Kafka production guardrails, DLQs, or schema enforcement: use kafka-resilience-and-schema-evolution
   • Live lag, Spark plans, or run-state diagnosis via MCP: use mcp-data-observability-integration
   • Governance, lineage, privacy, or access changes: use lineage-pii-and-governance
   • PII, PCI, HIPAA, PHI, or audit-bound data handling: use data-security-compliance-and-regulated-data
   • region-specific compliance, sovereignty, or localization rules: use regional-data-compliance-and-sovereignty
   • sustainability or ESG reporting data products: use esg-and-sustainability-regulatory-reporting
   • Schema changes or breaking contracts: use schema-evolution-and-contract-migrations
   • Informatica, Talend, or legacy ETL modernization: use enterprise-etl-and-data-integration-modernization
   • mainframe data offload or modernization: use mainframe-modernization-and-data-offload
   • Platform-team ownership, golden paths, or support-boundary design: use data-platform-operating-model-and-service-ownership
   • quality tool selection, rule severity, or quality operating model: use data-quality-platforms-and-rule-management

2. Choose the execution surface.

   • Python pipeline or validation package: python-data-engineering-and-pipeline-packaging
   • Scala Spark or JVM stream job: scala-data-engineering-on-jvm-runtimes
   • Java connector or metadata service: java-data-engineering-and-integration-services
   • Warehouse/dbt work: warehouse-and-schema-design + dbt-and-analytics-engineering
   • Snowflake-native warehouse pipelines: snowflake-native-pipelines-and-governance
   • BigQuery or Dataform platform workflows: bigquery-and-dataform-platform-engineering
   • Platform-native governance on AWS, Azure, GCP, or Databricks: use the matching governance skill plus the matching platform preset
   • Spark or distributed processing: spark-and-distributed-processing
   • Pipeline scheduling, DAG behavior, or event-driven orchestration: airflow-and-workflow-orchestration
   • Streaming or messaging: streaming-and-messaging-systems
   • CDC and source replication: cdc-and-incremental-loading or debezium-and-kafka-connect-cdc
   • Lakehouse table work: lakehouse-table-format-engineering or delta-lake-and-medallion-architecture
   • Infra and environment setup: terraform-and-data-platform-infrastructure
   • Multi-jurisdiction governance and reporting: regional-data-compliance-and-sovereignty or esg-and-sustainability-regulatory-reporting
   • Enterprise ETL estates: enterprise-etl-and-data-integration-modernization

3. Choose the platform preset.

   • AWS-heavy platform: aws-data-engineering
   • Azure-heavy platform: azure-data-engineering
   • GCP-heavy platform: gcp-data-engineering
   • Databricks-first platform: databricks-lakehouse-engineering
   • Snowflake-centric warehouse platform: snowflake-modern-data-platform
   • Informatica-centered estate: informatica-data-integration
   • Talend-centered estate: talend-data-integration
   • Mixed platform or migration state: multi-cloud-hybrid-data-engineering
   • Apache-first stack: load the matching Apache preset for Spark, Flink, Airflow, Kafka, or Iceberg

4. Recommend the fastest bootstrap asset.

   • New pipeline: start with templates/source-contract.yaml or templates/dataset-contract.yaml
   • Metrics or semantic work: start with templates/metric-contract.yaml
   • Operational recovery or release work: start with templates/incident-runbook.md, templates/backfill-plan.yaml, or templates/release-gate-evidence.yaml
   • Schema migration work: start with templates/schema-change-plan.yaml
   • Greenfield lakehouse adoption: use starter-packs/aws-lakehouse-starter.yaml
   • Medallion delivery: use starter-packs/databricks-medallion-starter.yaml
   • Warehouse modeling: use starter-packs/warehouse-analytics-starter.yaml
   • Streaming reliability: use starter-packs/streaming-reliability-starter.yaml
   • Production reliability and failure recovery: use starter-packs/production-reliability-starter.yaml
   • Data-platform release design: use starter-packs/data-platform-cicd-release-starter.yaml
   • Resilience drills and recovery proof: use starter-packs/resiliency-testing-starter.yaml
   • Validation and security review: use starter-packs/validation-security-review-starter.yaml
   • Privacy or governance work: use starter-packs/privacy-governance-starter.yaml
   • Regional compliance or ESG reporting: use starter-packs/regional-compliance-and-esg-reporting-starter.yaml
   • Test data and lower environments: use starter-packs/test-data-lower-environments-starter.yaml
   • Enterprise ETL modernization: use starter-packs/enterprise-etl-modernization-starter.yaml

5. Recommend an example project when concrete context helps. Prefer runnable scaffolds when the team needs executable proof; use architecture blueprints for /spec and /plan only. See examples/README.md for the full type column.

   Runnable:

   • Warehouse/dbt: examples/dbt-warehouse-marts/
   • Databricks medallion: examples/databricks-delta-medallion/
   • AWS lakehouse: examples/aws-s3-glue-athena-iceberg/
   • Kafka and Flink: examples/kafka-flink-streaming/
   • AWS serverless Spark and MSK reliability: examples/aws-serverless-spark-msk-reliability/

   Blueprint:

   • API or SaaS ingestion: examples/api-saas-to-warehouse-ingestion/
   • Privacy workflows: examples/privacy-retention-deletion-workflow/
   • Platform cutover: examples/multi-cloud-warehouse-cutover/

6. Choose the safe next command.

   • unclear intent or missing contracts: /spec
   • approved scope but too much ambiguity in sequencing: /plan
   • implementation work with clear acceptance criteria: /build
   • quality, reconciliation, or release gating: /validate
   • reliability, governance, or cost review: /review
   • replay, repair, or cutover work: /backfill
   • deployment or publish readiness: /ship

7. Pull in supporting references only when needed.

   • ingestion: references/api-saas-ingestion-checklist.md
   • file feeds: references/file-ingestion-checklist.md
   • quality: references/data-quality-checklist.md
   • security and compliance: references/security-compliance-regulated-data-checklist.md
   • asset discovery and install surfaces: registry/assets.json
   • testcase design: references/data-validation-and-testcase-patterns.md
   • resiliency testing: references/data-resiliency-testing-patterns.md
   • disaster recovery: references/data-platform-dr-bcp-checklist.md
   • operating model: references/data-platform-operating-model-checklist.md
   • platform-native governance: references/platform-native-governance-patterns.md
   • platform security: references/data-platform-security-checklist.md
   • anti-pattern review: references/data-engineering-anti-patterns.md
   • ETL and ELT strategy: references/etl-elt-modernization-checklist.md
   • mainframe modernization: references/mainframe-modernization-checklist.md
   • progressive release: references/progressive-data-release-patterns.md
   • test data: references/test-data-preparation-checklist.md
   • lower environments: references/lower-environment-masking-checklist.md
   • enterprise ETL modernization: references/enterprise-etl-modernization-checklist.md
   • regional compliance: references/regional-compliance-and-data-sovereignty-checklist.md
   • ESG reporting: references/esg-and-sustainability-reporting-checklist.md
   • streaming architecture: references/streaming-architecture-patterns.md
   • serverless Spark reliability: references/spark-serverless-reliability-patterns.md
   • Kafka production guardrails: references/kafka-production-guardrails.md
   • MCP observability: references/mcp-data-observability-patterns.md
   • cloud-specific common architecture patterns: references/cloud-data-engineering-architecture-patterns.md
   • pipeline orchestration across schedulers and clouds: references/pipeline-orchestration-patterns.md
   • data-quality tool boundaries and severity design: references/data-quality-tooling-and-rule-management.md
   • reference selection: references/README.md
   • streaming: references/streaming-checklist.md
   • recovery: references/incident-recovery-checklist.md
   • schema change: references/schema-migration-checklist.md
   • observability: references/observability-and-sla-checklist.md

Common Rationalizations

Rationalization	Reality
"It is only a small pipeline tweak."	Small data changes still break dashboards, contracts, or downstream jobs. Classify the work correctly first.
"We already know the stack."	Stack knowledge does not replace contract, replay, quality, or publish decisions.
"We can skip the example because the work is custom."	Examples reduce guessing and make rollout patterns safer even when the domain differs.
"Backfill is just rerunning the job."	Replay work can amplify bad data and downstream impact if windowing and reconciliation are unclear.

Red Flags

• the agent starts coding before choosing a workflow skill
• stack-specific implementation starts without a matching preset
• a replay or schema change is planned without explicit validation gates
• the agent loads the full repository instead of the smallest useful bundle
• there is no named next command for the current task

Verification

• The task type has been mapped to the correct workflow skill
• A matching platform preset has been chosen where relevant
• A starter template, starter pack, or example has been recommended when useful
• The safest next command has been identified
• Quality, governance, replay, and rollback risks were not skipped during triage