When I am deep in the codebase building out cloud infrastructure and AI workflows, I run into the same frustrating challenge: AI coding assistants are incredibly powerful, but they have terrible short-term memory. They treat every new session as a blank slate, meaning I have to constantly re-explain my architectural decisions, framework preferences, and deployment targets.
This is where we need to shift from endless prompt engineering to Context-Driven Development. By utilizing tools like the Microsoft Skills repository and persistent agent memory, we can build AI assistants that actually learn how we work and compound in value over time.
The Microsoft Skills Repository
Rather than crossing your fingers and hoping the underlying model guesses the correct framework or SDK version, Context-Driven Development feeds the model specific, modular knowledge just-in-time. The Microsoft Skills repository is the backbone of this approach, currently offering over 174 skills, plugins, and hooks.
These skills are essentially standardized Markdown files (SKILL.md). Before the AI generates a single line of code, your IDE invisibly injects these skills into the system prompt, forcing the LLM to adhere to the exact Azure SDK rules or Entra ID parameters you actually need.
Adding these enterprise-grade skills to your workspace is as simple as running a few CLI commands:
Bash
# Add frontend UI standards and design review capabilities
npx skills add microsoft/skills --skill frontend-design-review
# Equip your agent with Microsoft Entra ID integration knowledge
npx skills add microsoft/skills --skill entra-agent-id
npx skills add microsoft/skills --skill entra-app-registration
# Pull in Azure infrastructure and governance patterns
npx skills add microsoft/skills --skill foundry-governance
npx skills add microsoft/skills --skill azure-prepare
The Continual Learning Hook
While skills provide domain knowledge, the Continual Learning hook is what gives your agent long-term memory. Hooks are just automated scripts that trigger at key moments in the agentic loop.
When you attach the Continual Learning hook, your agent extracts your specific context and coding preferences from the conversation and tucks them away into persistent storage. Whether you mandate snake_case for database schemas or prefer strict TypeScript interfaces, the agent retrieves this context in future sessions automatically.
Awesome Copilot: A Cross-Platform Ecosystem
Beyond the official Microsoft repository, the Awesome Copilot community has built a massive collection of custom agents, instructions, and hooks. The best part? This ecosystem is totally interoperable. These skills aren’t locked to GitHub Copilot; they work flawlessly across:
- Claude Code by Anthropic
- Google Antigravity
- Visual Studio Code via the Foundry Toolkit
Taking It Offline with Foundry Local
You don’t even need a cloud connection to harness this power. For stringent data privacy or when you are just hacking away offline, Foundry Local is an absolute game-changer. It operates as an on-device AI runtime built on the ONNX engine, exposing a standard OpenAI-compatible API.
When running a device equipped with a Qualcomm Snapdragon X Elite processor or similar, Foundry Local seamlessly offloads execution to the NPU. This means you can run capable small language models (SLMs) like Qwen or Ministral completely locally. Your agent gets the exact same context, Microsoft skills, and Continual Learning memory—with near-zero latency, and your codebase never leaves your machine.
Scaling to Production: Azure Container Apps & Bicep
When it is time to take my local agent out of the sandbox and deploy it for the 45 of us over at Advantive, I don’t want to rewrite everything. You can push that exact same agent directly to the Microsoft Foundry Agent Service running natively on Azure Container Apps.
Following Infrastructure as Code (IaC) best practices, we use an Azure Bicep template to define the container environment and wire up the Azure AI Foundry endpoints. Your agent scales to zero when idle, but instantly spins up with its memory and skills fully intact the second a request hits.
Codefragment
// main.bicep
param agentName string = 'advantive-foundry-agent'
param containerImage string = 'myregistry.azurecr.io/myagent:v1'
resource containerApp 'Microsoft.App/containerApps@2024-03-01' = {
name: agentName
location: resourceGroup().location
properties: {
environmentId: containerAppEnvironment.id
template: {
containers: [
{
name: 'agent-container'
image: containerImage
env: [
{
name: 'FOUNDRY_ENDPOINT'
value: foundryResource.properties.endpoint
}
]
}
]
}
}
}
Conclusion: By combining Context-Driven Development, persistent memory hooks, ONNX-accelerated local SLMs, and robust Azure deployments, you transform your AI coding assistant from a generic guessing engine into a highly specialized, context-aware builder.
Links:
https://awesome-copilot.github.com
https://microsoft.github.io/skills
Leave a Reply