The bridge - AGENTS.md and CLAUDE.md

Here’s the part worth the price of admission. If you’ve taken the Claude Code course or the Codex course on this site, you may already have a convention file for a repo - an AGENTS.md, or a CLAUDE.md. You spent real effort getting it right: the conventions, the don’ts, the architecture notes an agent needs to not embarrass you in review. The natural assumption walking into Copilot is that you start over with copilot-instructions.md and re-encode all of it in a tool-specific file.

You often don’t. Supported Copilot surfaces can read those files natively, but support and discovery differ by surface and version. The same AGENTS.md, CLAUDE.md, or GEMINI.md may transfer without rewriting when the current support matrix lists it; verify before making it the only source of a required rule.

Turn it on

Reading these files is gated by a setting per format, so you flip the one matching the file you have:

• AGENTS.md - enabled by chat.useAgentsMdFile. Copilot reads the AGENTS.md at your repo root and treats it as instructions, the same way it treats copilot-instructions.md.
• CLAUDE.md - enabled by the relevant current Copilot/VS Code setting where supported. Copilot looks in documented repository and personal locations; the exact paths are surface-specific.
• GEMINI.md - supported by current Copilot instruction discovery on several surfaces; check the support matrix before relying on it in an IDE.

So the file you wrote in another course may not need to move or change. Enable the relevant setting where required, then confirm the discovered files in that surface before assuming the conventions apply to every request.

Nested AGENTS.md - preview

There’s a preview capability for nested AGENTS.md files - a per-directory file that applies to its subtree, useful in a monorepo where each package has its own conventions - behind chat.useNestedAgentsMdFiles. It’s experimental; know it exists, but don’t build your setup on it yet. Path-scoped rules have a stable mechanism, covered in the next lesson.

Which file, when you’re choosing fresh

If you’re not arriving with an existing file, you’re picking a format rather than rescuing work. A practical way to choose:

• If your team standardizes on a tool-agnostic convention file - one AGENTS.md that every agent on the team reads - that’s the most portable choice, and it’s the one the ecosystem is converging on.
• If you’ve already invested in CLAUDE.md from another tool, keep it; Copilot reads it.
• If you only ever use Copilot in this repo, .github/copilot-instructions.md is the most broadly supported repository-wide choice, with surface-specific limitations.

These aren’t mutually exclusive in spirit, but resist scattering the same conventions across three files - pick the one your team will actually maintain, and keep it the single source of truth.

Why this matters beyond convenience

The deep point isn’t that you saved a copy-paste. It’s that context is repo-shaped, not tool-shaped. Your team’s conventions describe how your code should be written; they shouldn’t have to be re-authored every time you adopt a new agent. The agents that read a shared convention file let you write your context down once and carry it across tools - which is the entire thesis of context engineering: close the gap once, reuse it everywhere.

If you came here from one of the other courses, this is the moment the courses stop being separate tracks and start being one practice with different front ends. If you didn’t, the Claude Code course and the Codex course walk through authoring the same kind of file from the other side - and whatever you write there lands here for free.

Next, the inverse problem: one repo, two zones with different conventions. Path-specific rules with applyTo.