Copilot supports .instructions.md files that live in your repo and give it persistent context — coding standards, framework conventions, architectural constraints. Same idea as CLAUDE.md for Claude Code or AGENTS.md for Codex, but with per-topic scoping so you can have separate instruction files for different concerns.

The technique

Instead of repeating context in every prompt, codify it once as a file. Copilot loads relevant instruction files based on the workspace and files you’re editing. The awesome-copilot repo has 160+ community-maintained examples to start from or adapt.

Why it’s on the radar

This is the same pattern emerging across every AI coding tool — persistent, repo-level context that travels with the code. Copilot’s version is worth trialing because:

• It’s per-topic scoped (one file per concern), not one monolithic file
• The community library gives you a starting point instead of writing from scratch
• The pattern is transferable — good instruction files can be adapted for Claude Code or Codex

Currently trying: context-engineering, self-explanatory-code-commenting, playwright-typescript, nextjs, github-actions-ci-cd-best-practices.

What to figure out

• Which instruction files actually improve output vs just adding token noise
• How well they compose when multiple apply to the same file
• Whether Copilot’s per-topic scoping is better than a single CLAUDE.md approach, or if it leads to fragmentation
• How to keep instruction files from going stale as the codebase evolves

Links

• awesome-copilot instructions