Yeda AI Tips · #32

Stop Letting Copilot Search The Whole Internet

You type a question into Copilot Chat, and the answer comes back generic — technically correct, practically useless, clearly not written with your actual codebase in mind. The usual culprit isn't the model. It's the question. Without a scope, Copilot has to guess where to look for an answer, and by default that guess is broad: it goes looking across the internet for anything relevant, which means "less accurate because there's too much context" to sift through. Narrowing the search is a matter of one character.

The problem: an unscoped question is a vague answer

Ask Copilot Chat "how do I fix this bug" with no other context, and it has to decide: is this a general programming question? Something about your specific framework? Something about the actual file open in your editor? Left to guess, Copilot defaults to the broadest possible interpretation — searching generally rather than looking at your project's own code. The fix is to tell it explicitly where the answer lives, before you ask the question.

The fix: chat participants

Type @ in the Copilot Chat box and a list of participants pops up — each one a "domain expert" scoped to a specific source of truth. The three most useful for day-to-day scoping:

GitHub also documents @github, which pulls in GitHub-specific capabilities (searching issues, PRs, and other GitHub data) beyond what's in your local workspace. Copilot Chat can sometimes infer which participant you meant from a natural-language question without you typing the @ — but typing it explicitly is the reliable way to guarantee the scope you want, especially for questions that could plausibly go either way.

How to use it

Type @ in the chat input. A list of available participants appears — pick the one that matches what you're actually asking about, then type your question after it:

@workspace where is axios used?

That single prefix is the difference between an internet-wide guess and an answer grounded in the file that's actually open, or the codebase that's actually yours.

Slash commands: scoping for common one-off tasks

Chat participants scope where Copilot looks. Slash commands scope what it does, and they're fastest when you've already selected code. Highlight a block, open inline chat or the chat panel, and type / to see the list. The commands GitHub documents include:

Typing / in the chat box shows the full list available in your current context — inline chat, the chat panel, and different IDEs can expose slightly different command sets, so it's worth checking what's on offer rather than memorizing a fixed list.

Combining the two

The strongest pattern is participant + selection + slash command: highlight the function you're worried about, then ask /explain or /fix directly on it. You've now scoped the question three ways — to a specific participant if needed, to a specific code selection, and to a specific task — leaving Copilot almost no room to wander into a generic, internet-wide answer.

Pitfalls to avoid

Takeaway

A vague Copilot answer is almost always an unscoped question. Type @ and pick a participant — @workspace for your code, @vscode for the editor, @terminal for your shell — then reach for /explain, /fix, or /tests on selected code for common one-off tasks. Specific input, specific output.

Resources

Building an AI feature? Yeda AI designs, audits, and ships production LLM systems.

Talk to us · Read the blog