microsoft/skills-for-fabric

1. This repository is not a running application; it is a content-and-package repository for AI assistant instructions around Microsoft Fabric. Its “architecture” is a distribution system for reusable skills, prompts, and tool configuration, not a service or UI stack.

2. The top-level repository acts as the source and umbrella bundle. The README.md describes the public packaging model, while CHANGELOG.md, LICENSE, SECURITY.md, and governance files support release, legal, and security boundaries.

3. The main payload is organized into bundles:

- skills/ for the full Microsoft Fabric skill set

- plugins/fabric-skills/ as the packaged full bundle

- focused bundles under plugins/ such as fabric-authoring, fabric-consumption, fabric-operations, and powerbi-authoring

These bundles are the primary unit of installation and consumption by AI tools.

4. Each bundle follows the same internal pattern: skills/ contains individual task-oriented skills, common/ contains shared cross-skill guidance, and agents/ contains role-specific instruction sets. The bundle-level .github/plugin/plugin.json and .mcp.json files define how the bundle is surfaced to plugin systems and MCP-aware tools.

5. common/ is the shared architecture layer. It contains workload-neutral fundamentals such as CLI usage, core Fabric concepts, and workload-specific “core” guidance for Spark, SQLDW, Eventhouse, Eventstream, Dataflows, and item definitions. The nested common/notebook-authoring/ directory provides reusable notebook-specific patterns and constraints.

6. skills/ is the workflow layer. Each skill folder maps to a distinct operator use case, such as authoring, consumption, operations, migration, or report design. The repo covers major Fabric surfaces: Spark, SQL data warehouse, Eventhouse/KQL, Eventstream, Dataflows Gen2, semantic models, Power BI reports, Activator, Fabric IQ, and migration scenarios from Databricks, HDInsight, and Synapse.

7. The agents/ folders define persona-based entry points. Agent files such as FabricAdmin.agent.md, FabricAppDev.agent.md, FabricDataEngineer.agent.md, FabricIQ.agent.md, and FabricMigrationEngineer.agent.md give the assistant role-specific behavior layered on top of the shared skill content.

8. The plugins/ subtree is the packaging/distribution boundary. It mirrors the content into installable bundles with their own manifests and scoped skill sets:

- fabric-skills = full library

- fabric-authoring = creation and automation workflows

- fabric-consumption = read-only discovery and querying

- fabric-operations = diagnostics and performance

- powerbi-authoring = report and semantic model authoring

This split is important for understanding what a user gets when installing a specific bundle.

9. The mcp-setup/ and docs/mcp-servers-guide.md area is the live-tool integration boundary. Skills provide instruction; MCP servers provide actual runtime access to Fabric APIs and data sources. The setup scripts and config template show that live capability is optional and environment-dependent.

10. prompt_examples/ provides validation and onboarding examples, not runtime logic. These examples show the kinds of end-user requests the skill bundles are designed to answer, such as medallion architecture, report generation, and workspace documentation.

11. The repository is content-heavy and code-light. The key “interfaces” are markdown instructions, plugin manifests, and configuration files rather than source code modules or compiled artifacts. That means the architecture diagram should emphasize content flows and packaging relationships instead of application runtime components.

12. The end-to-end data flow is: author writes or updates skill/common/agent markdown → bundle manifests expose that content to Copilot CLI or compatible tools → user installs a focused or full bundle → assistant reads the appropriate skill/agent instructions → if enabled, MCP tools connect to Fabric services for live operations. The repo’s purpose is to make that flow repeatable across many Fabric workloads.