Customize Copilot for a bounded inventory task

Instructions influence responses; they do not enforce permissions. A custom agent defines a role/tool profile; its target metadata does not mean “run in the cloud.” Select execution through the actual session target control.

Lab briefing

[!NOTE] A file existing on disk is not proof that the host loaded it.

Concepts · First task · Evidence checklist · Reset

Learning objectives

• Use the smallest customization primitive for each need.
• Verify discovery from a disposable workspace root.
• Distinguish instructions, task prompts, skills and handoffs.
• Implement one inventory change without assuming customization guarantees correctness.

Before you start

Prepare 15-customization using the common setup. The bundled Node inventory fixture needs no package install and replaces a mandatory external Blazor/database project. A C# service/controller adaptation is optional. Read the concepts reference.

Concepts and use cases

Exercise scenario

The fixture lists inventory categories. Add an item lookup by exact SKU, retaining the category contract and returning copies. Do not add persistence, authentication, cloud deployment, or a whole CRUD application.

Task 1 - Establish the baseline and workspace boundary

node --test --test-concurrency=1 inventory.test.mjs

Open only the copied fixture as the workspace root. Capture the selected role, harness, model and permission state. A customization file buried in another workspace is not proof it loaded.

Task 2 - Create focused instructions

1. Create .github/copilot-instructions.md with a short project description: integer quantities, immutable return values, standard library only, focused tests.

2. Create a path-specific file:

   ---
   applyTo: "**/*.test.mjs"
   ---
   Use node:test and node:assert/strict. Test invalid and missing IDs.
   A passing test must call production code and assert observable behavior.
   

3. Inspect the current customization discovery UI/references for the chosen host.

4. Attach one test file and ask which guidance applies.

5. Do not create conflicting instructions to “test precedence”: VS Code does not promise a combination order. Resolve contradictions explicitly.

Task 3 - Compare a prompt and a skill

1. In Local, create a prompt using agent: ask and read/search tools:

   ---
   description: Review an inventory change against its acceptance criteria.
   agent: ask
   tools: [read, search]
   ---
   Inspect the selected change and tests. Report missing cases with source paths.
   Do not edit files or claim tests ran unless their output is available.
   

2. Invoke it and record how it differs from automatic instructions.

3. Create a check-evidence/SKILL.md with frontmatter name and description, followed by steps to select/run existing checks and report actual evidence.

4. Inspect skill discovery in the Copilot harness when available.

5. Do not expect the prompt file to run on Agent Host; do not copy its whole text into always-on instructions as a workaround.

Task 4 - Define a planner and a bounded handoff

Create .github/agents/planner.agent.md:

---
name: inventory-planner
description: Plan a small inventory change without editing source.
tools: [read, search]
handoffs:
  - label: Implement reviewed plan
    agent: agent
    prompt: Implement only the reviewed inventory slice and run its focused tests.
    send: false
---
Inspect current source and tests. State acceptance criteria, files, checks and
non-goals. Do not edit, execute commands, or invent validation results.

Inspect the profile and tools in the host. Tool names/aliases and handoffs differ across surfaces; use the documented picker rather than adding guessed tool names. Do not set target: cloud: that is not the documented execution mechanism.

Task 5 - Complete a reviewed change

1. Ask the planner to specify findBySku(sku) with found, missing, invalid and copy-isolation cases.
2. Review the plan and the prefilled handoff prompt before sending.
3. Let Agent implement only that feature and its tests.
4. Review in a clean read/search-only session.
5. Run the full small fixture test file yourself.
6. Compare actual compliant behavior with the mere existence of instruction files.

Verify your work

• Customizations are discovered from the copied root.
• Instructions and prompts are not conflated.
• The planner has no edit/execute tools.
• Handoff is human-reviewed; unavailable features are recorded.
• New item lookup and old category behavior are both tested.
• No customization is described as a sandbox or deterministic policy.

Troubleshooting

If the profile is absent, inspect workspace root, filename and frontmatter. If a tool is unavailable, choose a supported alias rather than granting all tools. If the implementation ignores a convention, inspect context and tests; do not claim instructions enforce it automatically.

Independent practice

Adapt the same conventions to a C# inventory service and test project. Explain which rules are cross-stack and which applyTo patterns/tool names must change.

Reset

Save redacted discovery evidence. Close the copied workspace and its agent sessions. Restore only exercise customizations/source in that copy, never the curriculum’s shared .github directory.

Official references

• Custom instructions
• Prompt files
• Agent Skills
• Custom agents and handoffs
• Choose a harness