GitHub has published a practical guide for making Copilot understand a repository instead of treating every coding request as a blank slate. The approach puts project-specific context next to the code, where developers can review and maintain it.
The guide covers custom instructions and reusable skills, which can encode conventions and common workflows for a project. It also introduces custom agents for more specialized tasks. These mechanisms are meant to reduce the repeated prompting developers otherwise need when a coding assistant enters an unfamiliar codebase.
GitHub also shows how to connect Model Context Protocol, or MCP, servers such as Playwright. MCP is a standard way for an AI agent to access external tools and data. A browser-testing connection, for example, can give an agent a way to inspect an application rather than relying only on source files.
This is guidance rather than a new model release, and the results will still depend on the quality of the repository instructions and tools. The useful change is operational: teams can make their coding rules explicit and reusable, rather than expecting Copilot to infer them on every task.