GitHub outlines four methods for tailoring Copilot AI assistant to project conventions, including custom instructions, agent skills, custom agents, and MCP server connections.

GitHub published a tutorial on October 12, 2026, explaining how developers can configure the Copilot AI assistant to follow project-specific conventions and access external tools. The guide, authored by Senior Developer Advocate Christopher Harrison, describes four customization methods intended to reduce repetitive explanations of codebase standards.

According to the GitHub blog post, the company documented repository custom instructions, agent skills, custom agents, and Model Context Protocol (MCP) servers as ways to tailor Copilot behavior. Repository custom instructions use a .github/copilot-instructions.md file to define architecture patterns, naming conventions, and testing expectations that apply across a codebase. Path-specific instruction files can be placed in .github/instructions with an applyTo field to limit guidance to matching file types.

Agent skills package repeatable procedures in .github/skills subfolders. Harrison's example describes a quality-checks skill containing a SKILL.md file that documents which scripts run tests and linting, what each script validates, and how to troubleshoot common failures. The description helps Copilot determine when to load the skill.

Custom agents are positioned for tasks spanning multiple files, such as accessibility reviews. Repository-level agent profiles live in .github/agents and define the agent's purpose, instructions, tools, and project context, according to the post.

MCP servers connect Copilot to external tools, data sources, and services beyond the repository. Harrison cited Playwright as an example: after configuration, developers can instruct Copilot to use the Playwright MCP server to test functionality manually, allowing the assistant to open the application in a browser and interact with the interface.

Source link