GitHub Copilot just got a significant upgrade: Agent Skills. Announced December 18th, this feature lets you teach Copilot how you want things done—not through generic prompts, but through structured, reusable instructions that live alongside your code.
Building content agents at Windows scale, I’ve seen teams struggle with consistency. Same task, different approaches every time, because agents lack procedural knowledge. Skills address this directly.
What Are Agent Skills?
Think of a Skill as a playbook for your AI assistant. It’s a structured set of instructions (plus optional scripts or templates) that tells Copilot how to perform a specific task consistently.
Examples:
- “Refactor code following our team’s standards”
- “Generate release notes in our preferred format”
- “Review PRs according to our checklist”
Skills are version-controlled, shareable, and reviewable—just like code. They live in your repository or personal directory, making them discoverable and reusable across your team.
How Skills Fit the Ecosystem
This is where it gets interesting. We now have four complementary capabilities:
Custom Instructions
These tweak Copilot’s personality or tone. Example: “Always explain in detail” or “Use concise responses.” They don’t add new capabilities—they adjust how Copilot communicates.
Custom Agents
Specialized AI assistants with distinct personas and roles. Think of them as domain experts—a WinForms specialist, a security reviewer, a documentation expert. Custom agents (stored in .github/agents/) orchestrate complex, multi-step workflows tailored to your team’s standards. They’re activated for specific conversations and can be scoped repository-wide, organization-wide, or enterprise-wide.
MCP Servers
The Model Context Protocol exposes tools and resources: APIs, databases, file systems. MCP servers give Copilot access to external capabilities.
Agent Skills
Skills define how to use those tools in a workflow. They’re procedural knowledge—reusable modules (stored in .github/skills/) containing instructions, scripts, and resources. Skills are loaded automatically when contextually relevant, teaching Copilot specialized, repeatable tasks.
The analogy that works: MCP = toolbox. Skills = recipe book. Custom Agents = specialized chefs. Custom Instructions = personal style.
graph TD
A[Custom Instructions<br/>Personality & Tone] --> E[GitHub Copilot]
B[Custom Agents<br/>Specialized Roles] --> E
C[MCP Servers<br/>Tools & Resources] --> E
D[Agent Skills<br/>Procedures & Workflows] --> E
E --> F[Consistent, Context-Aware Output]
style A fill:#1e3a8a,stroke:#1e40af,color:#fff
style B fill:#4338ca,stroke:#4f46e5,color:#fff
style C fill:#0f766e,stroke:#0d9488,color:#fff
style D fill:#047857,stroke:#059669,color:#fff
style E fill:#334155,stroke:#475569,color:#fff
style F fill:#0e7490,stroke:#0891b2,color:#fff
The Power Move: Skills + MCP
Here’s where this matters for enabling agentic workflows: Skills and MCP servers complement each other.
Say you’re building an agent that handles release management. An MCP server gives Copilot access to your git repository, CI/CD system, and issue tracker. A Skill defines the workflow:
flowchart LR
A[MCP Server<br/>Git + CI/CD Access] --> B[Agent Skill<br/>Release Workflow]
B --> C1[Check unreleased<br/>commits]
C1 --> C2[Group by type<br/>features/fixes/breaking]
C2 --> C3[Generate release<br/>notes]
C3 --> C4[Create PR with<br/>release branch]
C4 --> C5[Tag relevant<br/>reviewers]
C5 --> D[Consistent Release<br/>Process]
style A fill:#0f766e,stroke:#0d9488,color:#fff
style B fill:#047857,stroke:#059669,color:#fff
style C1 fill:#1e3a8a,stroke:#1e40af,color:#fff
style C2 fill:#1e3a8a,stroke:#1e40af,color:#fff
style C3 fill:#1e3a8a,stroke:#1e40af,color:#fff
style C4 fill:#1e3a8a,stroke:#1e40af,color:#fff
style C5 fill:#1e3a8a,stroke:#1e40af,color:#fff
style D fill:#0e7490,stroke:#0891b2,color:#fff
Without the Skill, Copilot has the tools but not the playbook. With the Skill, Copilot follows your team’s process consistently.
Practical Implications
Based on my experience building agent-based workflows, here’s what Skills enable:
Codified Team Knowledge
Your team’s processes—code review standards, testing patterns, documentation practices—become executable instructions instead of tribal knowledge.
Workflow Consistency
The same task gets done the same way, every time. No more “I forgot the checklist” or “I didn’t know we had a standard for this.”
Onboarding Acceleration
New team members inherit experienced patterns immediately. Skills capture what senior developers would have explained manually.
Cross-Team Sharing
Organizations can build libraries of Skills for common workflows. Code review? There’s a Skill. API design? There’s a Skill.
Where This Is Headed
Skills rolling out now, with full availability by January 2026. The timing is notable—this lands right as MCP adoption accelerates and teams figure out how to structure agent-based workflows.
The combination matters: MCP provides the means, Skills provide the method. Together, they turn Copilot from a code autocomplete tool into something closer to a teammate who knows your team’s playbook.
Getting Started
If you’re planning to use Skills:
- Identify repetitive workflows - What does your team do the same way every time?
- Start small - Pick one well-defined workflow and codify it as a Skill
- Combine with MCP - If your workflow needs external tools, add MCP servers
- Iterate based on usage - Skills should evolve as your practices evolve
The official resources:
The Bigger Picture
This is part of the broader shift toward enabling agentic workflows. We’re moving from “AI that suggests code” to “AI that follows your team’s processes.”
Skills, custom agents, MCP servers, custom instructions—these aren’t competing approaches. They’re complementary layers:
- Instructions define personality
- Custom Agents orchestrate complex workflows with specialized roles
- MCP provides capabilities
- Skills encode reusable procedures
The teams that figure out how to combine all four will have AI assistants that actually work the way they work.
What workflows are you planning to codify first? Connect with me on LinkedIn to share your approach—I’m curious what patterns teams prioritize.