This repository is a gentle introduction to customizing GitHub Copilot. It uses a single, relatable scenario — reviewing a pull request before merge — to show how Copilot’s four customization options build on each other. By following along in VS Code, you’ll see how each layer changes the quality and consistency of Copilot’s responses.
Follow along in the companion repository: katiem0/customize-copilot-demo.
What You’ll Learn
- The four ways you can customize Copilot and what each one is for
- How layering customization changes Copilot’s output from generic to team-aligned
- How a custom agent shifts Copilot’s perspective on the same task
- Practical habits for getting consistent, repeatable results
The Four Customization Layers
The heart of the repository is a set of examples showing each customization type working toward the same goal. Together they demonstrate how you can encode your team’s expectations so Copilot follows them automatically.
| Layer | What it does |
|---|---|
| Instructions | Always-on guidance that sets baseline coding, testing, security, and accessibility expectations for every response. |
| Prompts | Reusable task templates that give a repeatable structure and output format for common work like reviews. |
| Skills | Portable playbooks and checklists Copilot can draw on when a task calls for them. |
| Custom Agents | Specialized roles that give Copilot a focused perspective — for example, reviewing as a tech lead. |
What the Demos Show
The repository includes a progression of short demos built around the same PR review scenario. Rather than prescribing exact prompts, they’re designed to let you feel the difference each layer makes:
- Starting simple — a basic, codebase-aware review shows what Copilot does with little guidance: useful but general feedback, with inconsistent structure and no clear recommendation.
- Adding the full layer — combining instructions, a prompt, a skill, and a custom agent produces a consistent, well-structured review that reflects your standards and ends with a clear merge decision.
- Isolating the agent — running the same task with and without a custom agent shows that the prompt structures the work, while the agent shapes the judgment and prioritization.
What to Compare
As you work through the demos, watch how the output changes as customization is added:
| With little customization | With the full customization layer |
|---|---|
| Useful but generic feedback | Feedback aligned to your team’s standards |
| Output format varies each time | Consistent, repeatable structure |
| May miss team-specific priorities | Applies your coding, testing, security, and accessibility expectations |
| Identifies issues without ranking them | Prioritizes risks and gives a clear recommendation |
| Reviews like a general assistant | Reviews from a specific, chosen perspective |
Getting the Most From It
- Start a new Copilot Chat for each demo — Chat uses conversation history as context, so a clean chat makes the effect of each customization easier to see.
- Read through the example instruction, prompt, skill, and agent files to see how each is written, then adapt them to your own projects.
Key Takeaway
Instructions define the rules. Prompts define the task. Skills define the process. Custom agents define the perspective.
Layering these together is what turns Copilot from a helpful generalist into a reviewer that reflects how your team works.