Agent skills for data engineering
Data Artisan
Ask a coding agent for a database schema and you get working code. You often do not get the indexes, the partition strategy, or the data quality checks, because nothing in the request said the thing has to survive contact with a terabyte. Data Artisan is a set of skill files meant to supply that missing context, one skill at a time rather than all at once.
One of nine planned skills shipped, MIT licensed, CI green
- of 9 planned skills shipped: schema design
- 1
- licensed and portable across agents
- MIT
The problem
Generated data code tends to be correct and naive at the same time. It runs, it passes a smoke test, and it is missing indexes on high-cardinality columns, has no partitioning story, no data quality checks, and no lineage or observability. Designs that are fine at a gigabyte fall over at a terabyte.
The gap is not the model's ability to write SQL. It is that nobody told it which patterns separate a working schema from a production one.
How it works
The one shipped skill is a single portable file describing schema design patterns and the reasoning behind them, in a format agents can load directly: indexing and partitioning strategy, constraints as defense rather than app-level validation, and per-database guidance across Postgres, Snowflake and BigQuery.
Because the file is plain and portable, it works in several agent tools (Claude Code, Cursor, Codex) rather than being tied to one vendor, and it can be pasted directly if the tooling is not supported.
Eight more skills are on the roadmap and not yet built: an ETL pattern library, a data quality framework, DuckDB and analytics, a local AI data stack, Kafka and streaming, advanced dbt patterns, data governance, and cloud cost optimisation. The README lists them as a roadmap, not a feature list, on purpose.
Where it landed
The repository is published under MIT with a markdown-lint CI workflow running against it. One skill, schema design, installs through the agent skills CLI or by copying the file; the other eight are documented as planned, not implied as shipped.
What it taught me
Writing the patterns down exposed the ones I was improvising
Turning working knowledge into instructions another system has to follow is a good way to discover which parts were habit rather than reasoning. The one skill that exists got substantially clearer once I had to justify each rule instead of just applying it.
Built with
- SQL
- PostgreSQL
- dbt
- DuckDB
- Agent Skills
- Markdown