Copilot concepts for hands-on work

Five questions before a task

The model is another choice, not a synonym for any row. The old screenshots shipped with the imported material show earlier UI layouts. Use the current Command Palette and the documentation below; do not infer current feature status, model access, or settings defaults from those images.

Choose the smallest useful workflow

1. Ask: request a trace of the real code with paths and uncertainty.
2. Plan: ask for acceptance criteria, affected files, risks, validation, and a stop condition.
3. Agent: implement one approved slice, then inspect its diff and actual tool output.
4. Review: compare the result with the acceptance criteria in a clean context.

Planning is useful for small risky changes too. Agent may perform planning internally; the three roles are not three mandatory product buttons in every CLI or harness.

Example: a library search feature

Ask: Trace the console action to the repository. Cite the files you inspected.
Separate what the code already does from what the requirement requests. Do not edit.

Plan: Add case-insensitive title search. Preserve existing loan and return behavior.
List the contract, empty/missing input cases, changed files, and focused test command.
Do not assume a whole book is unavailable because one physical copy is on loan.

Agent: Implement only the reviewed search slice in this disposable workspace.
Run the specified tests, report exit codes, and stop after two unsuccessful repairs
to explain the blocker. Do not change unrelated projects or publish a branch.

Context is selected evidence, not the entire repository

Attach the smallest useful file or selection, but also follow relevant interfaces, callers, tests, and data. #codebase requests retrieval; it does not guarantee every file is placed into model context. Closing a tab does not prove the file is inaccessible.

Instruction files influence responses. They do not enforce filesystem restrictions, guarantee code quality, or change the permission model. Resolve conflicting instruction files directly: the VS Code documentation does not promise a precedence order when it combines them. Inline completions are not the same instruction-aware chat workflow.

Customization choices

[!WARNING] VS Code hooks and agent-scoped hooks are Preview. Nested AGENTS.md support and extra Plan-agent tools are experimental where documented. Availability is affected by organization policy.

[!IMPORTANT] VS Code Agent Host sessions do not use prompt files. For the Spec Kit slash commands layout, use a supported Local session. The pinned Spec Kit v1.0.4 integration instead generates skills by default; see the versioned reference. Do not assume copying an old prompt into a skill preserves its scripts or arguments.

Features are not entitlements

Do not change a private repository to public just to complete a lab. Do not hardcode plan quotas, request counts, or a particular model as universally available. Record the capabilities actually available to your account. Local/offline exercise evidence must be labeled separately from cloud or live SDK execution.

Source register

Checked on 2026-09-06. These are product references, not copies of chat output.

Related guidance

• Environment and resource limits
• Harness guide
• Feature-status matrix