Copilot Features

Skill Instructions

Overview

GitHub Copilot provides a set of customization features that tailor how the AI assistant behaves. These features let you define coding standards, create reusable workflows, build specialized tools, and control how guidance is applied across your codebase. Each feature serves a distinct purpose—from always-on workspace rules to on-demand skills and lifecycle automation—and choosing the right one depends on your specific needs. Some features (Agent Skills and Agent Hooks) follow cross-tool standards and work across VS Code, Copilot CLI, and Claude Code; others are VS Code-specific.

This skill helps you:

• Understand each feature type and when to use it
• Select the correct feature based on your requirements
• Format and structure each feature type correctly
• Navigate detailed documentation for implementation

How to Use this Skill

CRITICAL: The current file is just an overview, but the files linked have much more information. Once a feature type is selected, you MUST read the associated linked file for much more information.

Use the Feature Reference for quick lookup, the Feature Selection Guide to make informed choices, and the linked documentation for comprehensive implementation guidance.

See also the Global Rules section for important rules which apply when creating or editing all features.

Feature Reference

Locations are configurable. Each feature type has a default workspace location shown above, but VS Code also supports user-level locations (e.g., ~/.copilot/skills/) and configurable search paths via chat.*Locations settings. See each reference file for the full list of supported locations.

Feature Purposes

• Custom Instructions – Define workspace-wide rules, standards, and conventions that are automatically applied to every chat request. Use for guidance that affects all development work.
• Instructions Files – Define scoped instructions for specific files, languages, or folders that activate based on file patterns or semantic task matching. Use when rules should apply only to certain contexts. (VS Code only)
• AGENTS.md – Provide custom instructions intended for all agents in the workspace, designed to be tool-agnostic and work across multiple AI assistants (GitHub Copilot, Claude Code, etc.).
• Agent Skills – Provide specialized, reusable capabilities with optional bundled assets (scripts, templates, checklists) that load on demand. Skills follow the Agent Skills open standard and are portable across VS Code, Copilot CLI, and the Copilot coding agent. Prefer skills over prompt files for new work due to broader tool support.
• Prompt Files – Provide explicit, reusable prompts for repeatable tasks that users invoke through slash commands or the command palette. Useful for lightweight, single-task prompts in VS Code. For portable or multi-file workflows, prefer Agent Skills instead. (VS Code only)
• Custom Agents – Define specialist assistants with explicit scope, tool restrictions, and constraints for role-specific workflows like planning, code review, or research. Supports both VS Code and the Copilot coding agent (via the target field).
• Agent Hooks – Execute custom shell commands deterministically at specific agent lifecycle points (SessionStart, PreToolUse, PostToolUse, etc.) to enforce security policies, automate quality checks, create audit trails, or inject context. Hook entry schema is shared across Copilot and Claude Code; container file formats differ.

Cross-Tool Compatibility

Agent Skills are the most portable feature type. When a feature could be implemented as either a skill or a prompt, prefer a skill for cross-tool compatibility.

Feature Selection Guide

When to Use Each Feature

Use this flow to quickly choose the right feature type:

Important: Only one slash command can be used at a time. This includes skills and prompt files invoked via slash commands, so you cannot manually select multiple skills/prompts in a single request.

Comparing Across Features

Use these criteria when choosing between similar feature types:

Instructions Files vs Agent Skills

• Does this apply to most work in a domain? → Instructions Files
• Does this apply to specific tasks or workflows? → Agent Skills
• Example: "Always use type hints in Python" → Instructions. "Generate Jest tests with mocks" → Skill.

Agent Skills vs Prompt Files

• Multi-step workflow with bundled assets (scripts, templates)? → Agent Skills
• Single focused task with parameterized inputs, VS Code only? → Prompt Files
• Needs portability across tools (CLI, coding agent, Claude Code)? → Agent Skills
• Example: "Security audit with checklist + scripts" → Skill. "Generate API route from method + path" → Prompt.

Agent Skills vs Custom Agents

• Same capabilities needed for all steps? → Agent Skills
• Need context isolation (subagent returns single output)? → Custom Agents
• Need different tool restrictions per stage? → Custom Agents
• Example: "Generate tests following team standards" → Skill. "Plan feature (read-only) → Implement → Review (different tools each stage)" → Custom Agents.

Custom Instructions vs AGENTS.md

• GitHub Copilot-specific standards? → Custom instructions
• Tool-agnostic rules for multiple AI assistants? → AGENTS.md
• Both can coexist. Use custom instructions for Copilot-specific guidance and AGENTS.md for cross-tool rules. Do not duplicate the same rules in both files.

Instructions/Skills vs Agent Hooks

• Need to guide agent behavior? → Instructions or Skills
• Need guaranteed deterministic execution? → Agent Hooks
• Need to block operations before they execute? → Agent Hooks (PreToolUse)
• Need to run code at specific lifecycle points? → Agent Hooks
• Example: "Always validate inputs" → Instructions. "Run linter after every file edit" → Hook.

Global Rules

Naming

• Use lowercase, hyphen-separated names for variable-name features (skills, prompts, agents, instructions files).
• Skill and agent names: max 64 characters, alphanumeric and hyphens only, no consecutive hyphens, no leading/trailing hyphens.
• Description fields: max 1024 characters. Keep descriptions concise and specific.
• Avoid generic names (e.g., "testing", "instructions", "standards").
• Use descriptive, action-oriented names that indicate purpose or capability (e.g., jest-test-generation, generate-api-route, code-reviewer).
• Fixed-name features (custom instructions, AGENTS.md) have specific filenames; no variation needed.

Generation Commands

VS Code provides commands to scaffold features interactively:

• /create-skill — Create a new agent skill
• /create-agent — Create a new custom agent
• /create-prompt — Create a new prompt file
• /create-instruction — Create a new instructions file
• /create-hook — Create a new agent hook
• /init — Initialize workspace customizations (creates copilot-instructions.md)

Other Rules

• For Agent Skills, the description is the only text used for automatic activation. It must state capability and when to use it.
• For Prompt Files and Custom Agents, the description is UI metadata; make it specific and task-focused.
• For Custom Instructions and AGENTS.md, do not use YAML frontmatter.
• Include one concrete example per feature type.

For detailed guidance on phrasing rules, constraints, and instructions, reference the writing-ai-instructions skill.

Intent Files

Intent files (INTENT.md or <name>.intent.md) record the operator's goals and decision history for a feature. They enable feature regeneration, maintenance, and quality assessment over time.

• Create an intent file when the operator provides goals for a new or existing feature.
• When iterating on a feature, append to the intent file's Log section.
• Do NOT include intent files in a skill's frontmatter references list.
• See Intent Files for the format specification and agent behavior rules.

Feature Interaction & Precedence

When multiple features activate simultaneously with overlapping guidance:

• Most specific wins. Instructions files with applyTo targeting specific files override broader custom instructions or AGENTS.md guidance.
• Explicit over automatic. User-invoked features (prompt files, slash-command skills) take precedence over auto-applied features.
• Later context wins. When features at the same specificity level conflict, the most recently attached or loaded context generally takes precedence. Note: load order for auto-applied features is not user-controllable, so minimize conflicts rather than relying on ordering.
• Hooks are independent. Agent hooks execute deterministically regardless of other features. Hook output does not conflict with instruction-based guidance — hooks control execution, instructions control reasoning.
• Scope hierarchy. Features exist at three scope levels — organization, workspace, and user — which may overlap. Workspace features generally take precedence over user-level defaults, and organization-level features apply across all repositories.

Minimize conflicts by keeping each feature focused on a distinct concern. Do not duplicate the same rule across multiple features.

Validation Checklist

• Name is lowercase and hyphen-separated (skills, prompts, agents)
• File path matches feature type
• Required frontmatter fields are present per feature type
• Description uses "Use when..." pattern (for skills and on-demand instructions)
• No banned vague language in content ("appropriate", "reasonable", "consider", "might")
• Examples are included
• See feature-specific validation checklists in each reference file