data-engineering-skills

为什么要用

核心特性

7 dbt skills — create, debug, test, document, migrate, refactor, incremental
2 Snowflake skills — find expensive queries, optimize by text/id
Skills auto-activate based on task context
Benchmarked: 53% real-world dbt pass rate, 84% Snowflake optimization pass rate
Includes dependency tracking and output verification steps

实时演示

实际使用效果

data-engineering-skill.replay ▶ 就绪

0/0

安装

作者仓库 · Claude Skills 官方文档

使用场景

实战用法： data-engineering-skills

Debug a failing dbt model without thrashing

👤 Analytics engineers facing a red CI run ⏱ ~20 min intermediate

何时使用： dbt run just failed with a cryptic error and you don't know if it's schema, lineage, or SQL.

步骤

Feed Claude the error + model

Use debugging-dbt-errors. Here's the stderr and models/marts/fct_orders.sql. Diagnose the root cause — don't guess.✓ 已复制

→ Claude reads upstream refs, diagnoses in order: schema → lineage → SQL
Apply the fix and verify

Apply the fix and run dbt build --select fct_orders+. Show me the before/after row counts.✓ 已复制

→ Clean run + row count verification

结果： Green CI plus a note of the root cause so it doesn't recur.

注意事项

Fixing a symptom downstream when the bug is upstream — The skill enforces an upstream-first diagnosis; don't skip the lineage step

搭配使用： bigquery-server · github

Find and fix your top expensive Snowflake queries

👤 Analytics leads with a climbing Snowflake bill ⏱ ~60 min intermediate

何时使用： Finance flagged the Snowflake bill and you need to cut it without breaking dashboards.

前置条件

Snowflake role with ACCOUNT_USAGE access — ACCOUNTADMIN typically, or a dedicated cost role

步骤

Identify worst offenders

Use finding-expensive-queries to list the top 20 queries in the past 30 days by credit cost. Group by app/user.✓ 已复制

→ Ranked table with credits, runtime, warehouse
Optimize each top one

For the top offender, use optimizing-query-by-id <query_id>. Propose rewrites with estimated savings.✓ 已复制

→ Rewritten SQL + before/after explain plan
Validate and deploy

Run the rewrite in a test warehouse — confirm same row count and shape before we swap.✓ 已复制

→ Safe swap candidate

结果： A prioritized list of fixes with measurable $ savings.

注意事项

Rewrites change row count silently — Always diff before deploying — the skill enforces this

搭配使用： bigquery-server

Migrate a pile of stored procs into dbt models

👤 Teams moving off legacy SQL to dbt ⏱ ~90 min advanced

何时使用： You've inherited a warehouse of nested CTEs and want them as documented, tested dbt models.

步骤

Point the skill at the source SQL

Use migrating-sql-to-dbt. Here's proc_monthly_revenue.sql. Convert it to dbt models with refs, documentation, and at least 2 tests per model.✓ 已复制

→ One or more .sql files, schema.yml with docs and tests
Build and verify

dbt build the new models and compare row counts to the legacy output.✓ 已复制

→ Row counts match within tolerance

结果： Legacy logic lives as testable dbt models.

注意事项

Hidden side effects in the proc (UPDATEs) — The skill flags side effects — separate them out, don't blindly convert

搭配使用： github

Convert a slow full-refresh model to incremental

👤 Analytics engineers with long-running dbt runs ⏱ ~45 min advanced

何时使用： A daily model has grown too big for full refresh.

步骤

Analyze the model

Use developing-incremental-models on models/events.sql. Pick a strategy (merge / insert_overwrite / delete+insert) and justify.✓ 已复制

→ Strategy + unique_key + partition / cluster keys recommended
Implement and back-fill

Apply the incremental config; outline a safe back-fill plan.✓ 已复制

→ Model + back-fill steps

结果： Daily runs that finish in minutes, not hours.

注意事项

unique_key gets duplicates on late data — Use merge and test it

组合

与其他 MCP 搭配，撬动十倍杠杆

data-engineering-skill + bigquery-server

Apply the same optimize-by-id pattern to BigQuery expensive queries

Adapt finding-expensive-queries for BigQuery INFORMATION_SCHEMA.JOBS and list top 20.✓ 已复制

data-engineering-skill + github

Open a PR per migrated model so each is reviewable

For every migrated model, open a GitHub PR with dbt test output attached.✓ 已复制

工具

此 MCP 暴露的能力

工具	输入参数	何时调用	成本
creating-dbt-models	model spec	New model	0
debugging-dbt-errors	error log, model	CI or local run failed	0
testing-dbt-models	model	Untested model	0
documenting-dbt-models	model	Undocumented model	0
migrating-sql-to-dbt	legacy SQL	Legacy migration	0
refactoring-dbt-models	model	Hard-to-read model	0
developing-incremental-models	full-refresh model	Runtime too long	0
finding-expensive-queries	lookback window	Cost hunt	ACCOUNT_USAGE query
optimizing-query-text	SQL text	Know the SQL, not the id	0
optimizing-query-by-id	query_id	Have the id from the UI	1 explain

成本与限制

运行它的成本

API 配额: Snowflake queries cost credits like any other — ACCOUNT_USAGE reads are cheap
每次调用 Token 数: 5–15k per dbt skill invocation
费用: Free skill
提示: Run finding-expensive-queries once weekly, not on every session

安全

权限、密钥、影响范围

最小权限： dbt: read + write to your project Snowflake: ACCOUNT_USAGE for cost skills

凭据存储： dbt profiles.yml / Snowflake key-pair in env; the skill doesn't store secrets

数据出站： None from the skill directly

切勿授予： SYSADMIN to the Claude session unless absolutely needed

故障排查

常见错误与修复

dbt compile succeeds, run fails with column not found

Stale lineage — dbt deps + dbt clean + dbt build --select model+

finding-expensive-queries returns nothing

ACCOUNT_USAGE has ~45min delay; also confirm role has SNOWFLAKE.ACCOUNT_USAGE

验证： SHOW GRANTS TO ROLE <role>

替代方案