🚀 Complete Vibe-Coding Setup: 7 Agents, 17 Skills, 6 MCP Servers — and Your Copilot Knows Your Entire Architecture

[AndreasKarz]

🚀 Complete Vibe-Coding Setup: 7 Agents, 17 Skills, 6 MCP Servers — and Your Copilot Knows Your Entire Architecture

The Problem We All Have

You open GitHub Copilot in Agent Mode, ask it to add a property to your domain model — and it generates code that ignores your naming conventions, puts business logic in the API layer, and forgets your test patterns.

Copilot is brilliant. But without context, it's just guessing.

What I Built

An open-source, production-tested setup that turns GitHub Copilot from a generic code generator into a context-aware team member that knows your architecture, accesses your tools, and follows your conventions — automatically.

Repository: github.com/AndreasKarz/DotNET-Developer-AI-Setup

What's Inside

.github/
├── AGENTS.md                        # Main agent with trust boundary & delegation
├── instructions/                    # 4 layer-specific coding standards
│   ├── general.instructions.md      #   → Architecture, naming, dependency rules
│   ├── graphql.instructions.md      #   → HotChocolate resolver conventions
│   ├── worker.instructions.md       #   → MassTransit consumer patterns
│   └── tests.instructions.md        #   → xUnit + Moq + FluentAssertions standards
├── agents/                          # 7 specialist expert agents
│   ├── CSharpExpert.agent.md        #   → Design patterns, async, performance
│   ├── DebugExpert.agent.md         #   → Build errors, runtime exceptions
│   ├── DevOpsExpert.agent.md        #   → Pipelines, Docker, Kubernetes, Helm
│   ├── HotChocolateExpert.agent.md  #   → GraphQL schema design, self-learning
│   ├── MongoDBExpert.agent.md       #   → Schema, indexing, query optimization
│   ├── MSSQLExpert.agent.md         #   → Execution plans, stored procedures
│   └── APIStitchingExpert.agent.md  #   → Gateway routing, query delegation
├── skills/                          # 17 deep-knowledge skills
│   ├── backend-developer/           ├── code-reviewer/
│   ├── database-specialist/         ├── devops-specialist/
│   ├── service-scaffolder/          ├── hotchocolate-specialist/
│   ├── penetration-tester/          ├── email-template-developer/
│   ├── git-specialist/              ├── microsoft-docs/
│   ├── agent-creator/               ├── skill-creator/
│   ├── prompt-creator/              ├── docx/  ├── pdf/  ├── pptx/  ├── xlsx/
└── prompts/
    └── Implement_from_PBI.prompt.md  # PBI → Analysis → Plan → Code → Tests → Review

The Numbers

What	Count
Specialist Agents	7 — each with domain expertise and delegation rules
Skills	17 — from backend development to penetration testing
Instructions	4 — layer-specific, auto-loaded by file pattern
MCP Servers	6 — Azure DevOps, MongoDB, SQL Server, Microsoft Learn, Memory, Sequential Thinking
Lines of AI Configuration	~5,000+ — battle-tested on a production .NET backend

What Makes This Different

1. One Command Setup

npm run setup-mcp

Installs and configures all 6 MCP servers, warms the NPX cache, generates mcp.json for VS Code — done in under 2 minutes.

2. The Agent Knows Your Rules

Ask it to add a property and it will:

• Follow your naming conventions (from general.instructions.md)
• Use the correct repository pattern (from backend-developer skill)
• Expose it correctly in GraphQL (from graphql.instructions.md)
• Write tests matching your patterns (from tests.instructions.md)
• Skip N+1 queries by using DataLoaders (from hotchocolate-specialist skill)

3. Self-Learning Agents

The HotChocolate Expert and DevOps Expert fetch official documentation before answering — so they stay current even as frameworks evolve.

4. Security Built-In

A comprehensive trust boundary blocks prompt injection from 12+ untrusted sources (code comments, DB results, pipeline YAML, GraphQL schemas, images...). The penetration-tester skill runs OWASP-grade security assessments.

5. End-to-End Prompt Workflows

The Implement_from_PBI prompt takes a Work Item ID and drives the full cycle:
PBI Analysis → Codebase Analysis → Implementation Plan (you confirm) → Code → Tests → Code Review

6. Not Just .NET — Adaptable to Any Stack

The repo includes a CUSTOMIZATION.md with 5 complete migration guides:

• GraphQL → REST (ASP.NET Controllers + MediatR)
• Azure DevOps → Jira + Confluence
• C#/.NET → Java/Spring Boot
• MongoDB → PostgreSQL + EF Core
• MassTransit → RabbitMQ Direct

Each guide shows exactly which files to change and includes a ready-to-use Copilot prompt that rewrites the entire configuration.

7. MonoRepo Ready

Symlink the .github/ folder into any number of repositories — one source of truth, zero duplication.

Architecture Review: APPROVED

The entire setup has been through a senior architecture review covering all 4 instructions, 7 agents, 17 skills, 1 prompt, and the main AGENTS.md. Verdict: Production-Ready.

Key findings:

• Zero duplication of trust boundary (centralized, referenced everywhere)
• Clean agent-skill separation (agents orchestrate, skills implement)
• Token-efficient design (~200-250 lines per agent)
• Consistent cross-references across all delegation tables

Live Demo? I'd Love To! 🎤

I'm planning to present this setup at a GitHub Copilot Dev Day event on March 19, 2026 as a Community Session. The demo would show:

1. Running npm run setup-mcp from zero to fully configured agent
2. Live-coding a feature using the Implement_from_PBI prompt with Agent Mode
3. Specialist agent delegation in action (@MongoDBExpert, @DebugExpert)
4. How the same setup adapts to different tech stacks via CUSTOMIZATION.md

If you're organizing a Dev Day and looking for community session content — I'd be happy to contribute!

Try It Yourself

git clone https://github.com/AndreasKarz/DotNET-Developer-AI-Setup.git
cd DotNET-Developer-AI-Setup
npm run setup-mcp

Open VS Code, start Copilot Chat in Agent Mode, and ask:

"Review the coding standards that apply to test files in this project."

Watch it load the right instructions automatically. Then try:

"Break down the steps to add a new 'Notes' property to an existing entity."

Watch it use sequential thinking to produce a layered implementation plan.

Feedback, questions, and ideas for improvement are very welcome!