Rolling out an AI-native workflow across a data team
A shared skills marketplace, the rules that keep it clean, and a dbt skill set that lets real analytics engineering run through an agentic flow with its own tests and audits.
The problem
AI coding tools get adopted one person at a time. Everyone writes their own prompts, output quality depends on who ran it, nothing is reusable, and two people in the same repo collide. The productivity story looks convincing in a demo and then evaporates at team scale. The ceiling shows up fastest in analytics engineering, where the work is too intricate to hand to an agent that has no way to check itself.
What we built
We designed and implemented the architecture for rolling Claude Code out across an entire data team: a shared skills marketplace so repeated work is packaged once and used by everyone, meta-skills and meta-rules governing how skills are written and reviewed so quality holds as the library grows, and the collaboration infrastructure for several people to build and ship in parallel without stepping on each other.
The deepest part of the library was dbt. We built a set of skills for handling, managing, and scaling dbt so that genuinely complex analytics engineering could be executed by an agentic flow rather than by hand, and built the tests and audits into those same skills so the output is verified rather than assumed. That last part is what makes the agentic flow usable on production models instead of a demo. Alongside it we built targeted skills for the team's BI platform, covering Omni YAML management and branch management, so semantic model changes are version-controlled, reviewable, and safe to run concurrently.
The impact
Analytics engineering work that would normally queue behind a person could be run through the agentic flow and checked automatically. Migration work on the team ran materially faster for the same reason: the tooling absorbed the mechanical parts. The more durable result is that the skills library compounds, since each project adds to it rather than starting from a blank prompt.
What they own now
The marketplace, the skills, the dbt tests and audits, and the meta-rules that govern all of it, in their own repositories.
We publish client work without naming the client. No logos, no business metrics, no data. If you want references, we make them on a call, with the client's blessing.
Same build, your data. Tell us what is broken and we will tell you what it takes.
Book an intro call