Part 6: Specialized Agents & Custom Instructions

Your app is built, themed, and polished. Until now Copilot has worked as a generalist — it does whatever your current prompt asks. In this part you’ll give it specialized personas and always-on project rules so the same prompt produces sharper, more consistent results.

You’ll learn three layers of customization:

Built-in agents — the specialists Copilot already ships with
Custom instructions — project rules Copilot loads automatically
Custom agents — your own reusable specialists defined in .agent.md files

Section 1: Meet the Built-in Agents

You’ve already been using agents without knowing it. /plan and /review are built-in agents — specialized modes with their own focus and tools.

Agent	How to invoke	What it does
Plan	`/plan` (or Shift+Tab to cycle modes)	Creates a step-by-step plan before coding
Code review	`/review`	Reviews staged/unstaged changes for bugs and quality
Rubber duck	`/rubber-duck`	Gives an independent critique of your current approach
Security review	`/security-review`	Scans changes for security vulnerabilities
Explore	automatic	Investigates the codebase when a prompt needs research
Task	automatic	Runs builds, tests, and installs, then reports back cleanly

Try a couple you haven’t used yet, right on the Mona Mayhem code:

/rubber-duck Is my battle fetch logic in src/pages/index.astro resilient if one user request fails but the other succeeds?

/security-review

💡 The Explore and Task agents run automatically. When you ask “how is contribution data loaded?” Copilot silently uses Explore; when you ask it to “run the build” it uses Task and reports a clean pass/fail summary.

Result: You now know the “hidden” agents behind the commands you’ve been using — and two new quality gates (/rubber-duck, /security-review).

Section 2: Path-Specific Custom Instructions

In Part 1 you generated .github/copilot-instructions.md with /init. That file applies to every prompt. But some rules should only apply to certain files — Astro component conventions shouldn’t clutter a conversation about a TypeScript API route.

Path-specific instruction files solve this. They live in .github/instructions/ and use an applyTo glob in their frontmatter.

Task 1: Add Scoped Instructions

Ask Copilot to create scoped instruction files:

Create two instruction files in `.github/instructions/`:

1. `astro.instructions.md` - Applies to all Astro files (**/*.astro)
   - Should cover Astro component standards, TypeScript in scripts, neon arcade theme
   - Dark background (#0a0a1a), green (#5fed83), purple (#8a2be2) accents, Press Start 2P font
   - Keyboard accessibility requirements

2. `typescript.instructions.md` - Applies to all TypeScript files (**/*.ts)
   - Should cover error handling with try/catch blocks
   - Typed request/response interfaces for API routes
   - RESTful conventions for endpoints in src/pages/api/

After Copilot creates the instruction files, use /diff to review what was generated.

How applyTo works:

`applyTo` value	When it applies
`"*/.astro"`	Any Astro file
`"*/.ts"`	Any TypeScript file
`"src/pages/api/**"`	Anything in the API folder
(no frontmatter)	Every conversation — the default

💡 You can keep a short AGENTS.md at the repo root and pull in detail with @-imports, e.g. @.github/instructions/astro.instructions.md. Copilot expands the reference automatically.

Run /instructions to see which instruction files are currently loaded and toggle any on or off. Copilot CLI reads instructions from AGENTS.md, .github/copilot-instructions.md, and .github/instructions/**/*.instructions.md automatically.

You may have to run /restart to reload Copilot’s context and pick up the new files.

Task 2: See Them Work

Ask Copilot to add a small feature and watch it follow the scoped rules without being reminded:

Add a small "clear results" button to the battle page that resets both inputs and the results area.

The result should already use the neon colors, typed data, and keyboard accessibility — because the instructions applied automatically. Commit your instruction files.

Section 3: Create Custom Agents

Custom instructions shape every matching prompt. Custom agents go further — they’re reusable specialists you switch into on demand. Same prompt, expert output.

An agent is a Markdown file with a .agent.md extension: YAML frontmatter (metadata) plus a persona written in Markdown. Both VS Code and Copilot CLI read the same files from .github/agents/.

Task 1: Build a Retro UI Reviewer

Ask Copilot to create .github/agents/retro-ui-reviewer.agent.md:

Create a custom agent at `.github/agents/retro-ui-reviewer.agent.md` that specializes in reviewing UI for retro arcade themes. It should focus on neon aesthetics (dark #0a0a1a, green #5fed83, purple #8a2be2), animation quality, consistency, and how new elements match the arcade theme.

Task 2: Build an Accessibility Auditor

Ask Copilot to create .github/agents/a11y-auditor.agent.md:

Create a custom agent at `.github/agents/a11y-auditor.agent.md` that audits for WCAG 2.1 AA compliance, checking keyboard navigation, color contrast, ARIA labels, screen-reader announcements, and prefers-reduced-motion support.

After Copilot creates the agent files, use /diff to review what was generated.

Task 3: Put Your Specialists to Work

List and switch agents with /agent, or launch straight into one:

/agent a11y-auditor

Audit @src/pages/index.astro for accessibility issues.

To switch mid-session, run /agent again and pick another specialist (or no agent to return to the default experience). Agent selection is session-scoped — a new session starts fresh.

/agent retro-ui-reviewer

Review the battle page styling and flag anything that drifts from the arcade theme.

Task 4: Specialist vs Generic — See the Difference

Run the same prompt twice: once in the default experience, once with retro-ui-reviewer selected.

Add a small "rematch" button below the results.

The default answer works. The specialist version keeps the palette, adds hover/focus states, and matches the animation style — because the expertise is baked into the agent. Commit your agent files once you’re happy.

💡 YAML quick reference: description is the important field — it tells Copilot when the agent is useful. tools is optional (omit it to allow all tools); use aliases like read, edit, search, execute. The model property is honored in VS Code and ignored by the CLI, so cross-platform agent files are safe.

Check your understanding

When is a custom agent a better fit than custom instructions?

Check your answer

Custom instructions provide always-on guidance for work in their scope. A custom agent is an optional specialist with its own context, tools, and role, which makes it better for focused delegated work such as a security review or accessibility audit. The two can work together: instructions preserve project conventions while the agent supplies task-specific expertise.

Go deeper: Compare Copilot CLI customization features.

✅ Part 6 Complete

You’ve learned how to:

Recognize the built-in agents (/plan, /review, /rubber-duck, /security-review, Explore, Task)
Add path-specific instructions with applyTo so rules apply only where they belong
Build and switch between custom agents that carry your project’s expertise
Prove the value of specialists by comparing generic vs. agent-guided output

Next: In Part 7 you’ll package repeatable expertise as Skills that Copilot loads automatically when your prompt matches.