All case studies
Open SourceOpen source (MIT) · 2026

Nine skills that build a Power BI project

An open-source Agent Skills toolkit that goes from a plain-English request to a valid, branded PBIP.

Author and maintainer

The problem

Agent tooling for BI tends to stop at generating a SQL query or a picture of a chart. Neither is a deliverable. The actual unit of work in Power BI is a project (a semantic model, visuals, a theme), and nothing was generating one that would open without hand-repair.

What I did

Architecture

Nine Agent Skills composing into a Power BI Desktop ProjectENTRYPIPELINEPOSTorchestratorNLQ entry pointdata-source→ data-model.jsonsemantic-mapperstage 1 · TMDLvisual-selectorstage 2visual-generatorstage 3 · PBIRproject-packagerstage 4 · validatetheme-brandingbi-dash-creator121 · data-model confirmation gate2 · scaffold confirmation gate
An entry-point orchestrator, a source-agnostic connector, a four-stage build pipeline and two post-pipeline skills. Two human confirmation gates sit in the flow so the agent stops and asks rather than guessing a schema at scale.

What went wrong first

Composability is a contract problem

Nine skills only compose if each one's output is the next one's declared input. Writing the contracts first, and treating data-model.json as the pivot, was what stopped this becoming one unmaintainable mega-skill.

Agents guess confidently when schemas are ambiguous

Source introspection cannot always infer relationships. Rather than let the agent invent them, the connector emits an explicit clarification list and the orchestrator gates on it.

Results

composable skills
9composable skills
pipeline stages, independently testable
4pipeline stages, independently testable
licensed and public
MITlicensed and public

It closes the last mile that most natural-language BI tooling leaves open: not a chart image, but a project a BI developer can open, review and take ownership of.

Stack

PythonAgent SkillsTMDLPBIPPBIRDAXClaude CodeMCP
Repository