Skip to main content

MCP Server, Skills, and copilot.md Practice

This tutorial gives a practical path to three advanced Copilot concepts used in the workshop exercises repository:

  • MCP server configuration in VS Code
  • Reusable skill playbooks (SKILL.md)
  • Prompt drills in copilot.md

1. Prerequisites​

  • VS Code with GitHub Copilot enabled
  • Workshop exercises repository available locally
  • Node.js installed
cd workshop-exercises
npm install
npm run dev

2. MCP Server Setup (Local VS Code)​

The exercises repo already ships workshop-exercises/.vscode/mcp.json with two servers:

{
  "servers": {
    "playwright": {
      "command": "npx",
      "args": ["@playwright/mcp@latest"]
    },
    "github": {
      "type": "http",
      "url": "https://api.githubcopilot.com/mcp/"
    }
  }
}

playwright drives a real browser for the Exercise D E2E work; github gives Copilot Chat access to repo/PR/issue context. If your organization uses additional MCP servers, add them in the same file using the server settings provided by your platform team.

Verify MCP access​

@github and skill discovery are two separate things — the github MCP server gives Copilot Chat access to repo/PR/issue data, while skills are read locally from .github/skills/. Verify each independently.

Verify the github MCP server with a real repo-oriented prompt:

@github what are the open pull requests in this repo?

You should see a response reflecting real GitHub data, not a generic answer.

Verify skill discovery with a plain Copilot Chat prompt (no @github prefix):

what skills are available in this workspace?

You should see the three skills listed in the next section reflected in the response.

3. Skill Practice in the Exercises Repo​

The exercises repository includes three practice skills:

  • .github/skills/pact-contracts/SKILL.md
  • .github/skills/flaky-test-hunt/SKILL.md
  • .github/skills/test-generation/SKILL.md

Current GitHub Copilot guidance treats skills as reusable folders of instructions, scripts, and resources. In practice, you will most often see them in project-level locations such as .github/skills, .claude/skills, or .agents/skills, and in personal locations such as ~/.copilot/skills or ~/.agents/skills.

If you use GitHub CLI in environments where skill discovery is enabled, you can also explore installable skills with commands such as gh skill.

Pact contracts drill​

Despite the name, this skill isn't Pact consumer-driven-contract testing — it reviews API tests against the Express route/response-envelope contract in src/app.ts. Use this prompt:

Use the pact-contracts skill to inspect tests/api/auth-and-users.test.ts against src/app.ts.
Report missing assertions by severity and propose minimal test updates.

Flaky test hunt drill​

Use this prompt:

Use the flaky-test-hunt skill to diagnose the notification test marked FLAKY
in tests/unit/notificationService.test.ts. Return findings first, then a
deterministic fix.

Test generation drill​

Use this prompt:

Use the test-generation skill to scaffold tests for src/services/discountService.ts.
Reuse factories from tests/factories.ts where applicable.

4. copilot.md Prompt Workbook​

Open:

  • copilot.md

Complete at least these sections:

  1. Exercise A — Generate Unit Tests for calculateDiscount (weak vs. strong prompt)
  2. Exercise C — API Tests for the Checkout Pipeline (cart / full pipeline / error paths prompts)
  3. Exercise E — Context Engineering (generate with context / review for CI prompts)
  4. Prompt Iteration Drill (the comparison table at the end)

Goal: compare weak vs. specific prompts and record output quality differences.

5. Hands-on MCP Practice Drills​

Lightweight MCP-style investigation drills, no external services required. Work through at least the first.

Drill 1 — Investigate an API Failure​

  1. Intentionally break one assertion in tests/api/auth-and-users.test.ts.
  2. Run npm run test:api.
  3. Work the investigation loop: reproduce → localize → hypothesize → validate → fix.
  4. Restore the passing state.

Drill 2 — Contract Drift Detection​

  1. Temporarily change one response field in src/app.ts.
  2. Re-run the API tests.
  3. Decide whether to fix the implementation to preserve the contract, or update the tests because the contract intentionally changed.

Drill 3 — Auth Boundary Checks​

  1. Add one test for a missing Bearer token.
  2. Add one test for a malformed Bearer token.
  3. Add one test for the valid-token path.

MCP tooling mindset​

Treat each drill as tool-assisted investigation, not guesswork: read the exact failure output, inspect only the relevant files first, and prefer small, verifiable changes.

6. Validation Checklist​

  • MCP config file exists and is valid JSON
  • You successfully used at least one @github query in Chat
  • You completed at least one skill-guided review
  • You completed at least one copilot.md prompt drill
  • API tests still pass
npm run test:api

7. Agent Mode and Workspace-Style Practice​

Use this section to practice modern Copilot workflows where planning and implementation are separated.

A) Plan-first drill​

In Copilot Chat, switch to Plan mode and prompt:

Plan updates to tests/api/auth-and-users.test.ts to increase negative-path coverage for register/login.
Include status assertions, envelope assertions, and one sensitive-field check.
Do not edit files yet.

Review the plan before moving to edits.

B) Agent execution drill​

Switch to Agent mode and prompt:

Implement the approved test updates in tests/api/auth-and-users.test.ts.
Run only the API test command, report findings first, and keep changes minimal.

C) Cost-aware prompting drill​

Repeat the same task twice and compare quality/cost:

  • Broad prompt with no file context
  • Focused prompt with #file and explicit exit criteria

Capture what changed in output quality, speed, and number of follow-up edits.

8. Team Governance Notes (Conference Edition)​

For team and enterprise environments, align this workflow with your organization policy:

  • Approved model list and usage policy
  • Agent mode guardrails (who can run autonomous edits)
  • Review requirements before merge
  • Data handling policy for prompts and logs

Treat Copilot output as a draft until human-reviewed and validated in CI.

9. Common Pitfalls​

  • Adding overly broad prompts without file context
  • Treating generated output as final without review
  • Changing implementation and tests simultaneously without isolating root cause
  • Running full test suites when a focused suite is enough for diagnosis

Next Steps​

  • Return to the Copilot Overview workshop page for Copilot fundamentals
  • Continue to Exercise C — API Tests and apply the same skill/MCP workflow on API test evolution