agent-customization

Create a .instructions.md file in .github/instructions/ with a description in the YAML frontmatter and optional applyTo glob patterns. Use a keyword-rich description starting with "Use when..." so the agent can discover it on demand.

Both appear as slash commands in chat. A prompt file handles a single focused task with parameterized inputs, while a skill provides a multi-step workflow with bundled assets like scripts and reference documents.

Silent failures usually come from YAML frontmatter errors such as unescaped colons, tabs instead of spaces, or a name that does not match the folder name. Also verify the description contains the trigger phrases the agent uses for discovery.

Use hooks when behavior must be enforced deterministically, such as blocking dangerous commands or auto-formatting after edits. Instructions only guide agent behavior non-deterministically, while hooks run shell commands at lifecycle events like PreToolUse.

Use only one, not both. copilot-instructions.md in .github/ is recommended for project-wide standards, while AGENTS.md is an open standard supporting monorepo hierarchies with files in root or subfolders.

It loads the instruction into the context window on every interaction regardless of relevance, which burns context. Prefer specific globs like **/*.py or src/api/** unless the instruction truly applies to all files.