Building a Skills Library: Teaching AI Agents Your Org's Tribal Knowledge

AI coding assistants like GitHub Copilot are impressive at generating code. But ask one to "create a Jira story under the Platform epic with BDD acceptance criteria following our team's format" and it falls apart — because it doesn't know your org. Over the past year, I've built and maintained a skills library — a structured collection of reusable knowledge modules that transform a general-purpose AI assistant into a domain-aware engineering partner. Each "skill" encodes a specific capability: how to call our Jira API, our Azure DevOps patterns, our production support runbooks, our API design standards, even our commit message conventions. In this talk, I'll cover the architecture: the SKILL.md format (instructions + references + scripts + examples), how skills compose into agents (specialized personas with scoped tools), and how a single conversation can go from "implement this Jira ticket" to a pushed branch with tests — what we call the "one-shot" workflow. I'll demo three real skills live: a Jira CLI skill that creates governed work items from natural language, a spec-driven development workflow that generates implementation plans from ticket descriptions, and a production support skill that knows our Kubernetes namespaces, API endpoints, and escalation paths. We'll also cover the hard parts: preventing hallucinated API calls, enforcing guardrails (read-only in production), handling tool failures gracefully, and keeping skills maintainable as your systems evolve. This pattern isn't GitHub-specific — it works with any agentic AI system. You'll leave with a template to start encoding your own team's knowledge into reusable, composable AI skills.