GitHub Copilot: from inline completion to Copilot Workspace

GitHub Copilot is the default. It's also the most evolved — inline, chat, workspaces, and agent mode are all separate features with distinct strengths.

The feature matrix

• Inline autocomplete — the original, still the most-used. Ghost text as you type.
• Copilot Chat — sidebar conversation; can reference files, symbols, repo.
• Copilot Workspaces — larger, multi-file changes with a plan view.
• Copilot Agent Mode (or Copilot Coding Agent) — autonomous task execution, can open PRs.
• Copilot Review — AI-assisted PR review comments.

Each is useful for different sizes of work.

When to use which

Inline autocomplete that works

Inline Copilot is trained on your open files as context. Practical tips:

• Name the function well first; then let Copilot write the body. The name is the most important prompt.
• Write a one-line comment describing intent if the code isn't obvious from context.
• Keep related files open in your editor — they become context.
• Reject by pressing Esc quickly on bad suggestions; don't edit after accepting.

Copilot Chat — the underused one

Most engineers treat Chat as "autocomplete but fancier." It's more:

• Inline chat on a highlighted block — "explain this," "refactor this to use Zod," "add error handling."
• Chat with file references — @workspace, #file:src/auth.ts — explicitly scope context.
• Slash commands — /explain, /tests, /fix.

Workspaces

For tasks that touch 3-10 files:

1. Describe the change in plain language.
2. Copilot generates a plan.
3. You review and tweak the plan.
4. Copilot generates specs per file.
5. You commit or revise.

Workspaces are better than one-shot for deliberate, reviewable changes.

Agent Mode

The newest territory — Copilot's autonomous mode. Assigns itself an issue, creates a branch, makes changes, runs tests, opens a PR.

Use it for:

• Well-specified bug fixes.
• Small features with clear acceptance criteria.
• Chore work (dependency upgrades, test fixes).

It's not a replacement for engineering judgment on ambiguous or architectural tasks. Human review on the resulting PR remains non-negotiable.

Configuration that makes it better

• Custom instructions (Settings → Copilot → Custom instructions). Per-repo or per-user guidance: tone, style, project conventions. Biggest single quality improvement you can make.
• Content exclusions for sensitive files (secrets, proprietary configs).
• Model selection — choose between available models per query (Claude, GPT, etc.) for Chat. Different models suit different tasks.

The cost/quality knob

Copilot's autocomplete was tuned for speed; suggestions are usually fast and often right-ish. The Chat modes using frontier models produce noticeably better outputs for hard questions at slight latency cost.

When the simple suggestion isn't good enough, escalate to Chat or Workspaces rather than retrying autocomplete harder.

What it isn't

• A replacement for understanding your code. Copilot sometimes produces plausible-looking code that doesn't fit your actual architecture.
• Trustworthy for security-sensitive code. Authentication, crypto, input validation — review closely.
• Good with your internal libraries unless they appear in its training or you give it context.