Skip to content

GitHub Copilot Fundamentals

Encode what is always true in instructions, what is reusable and triggerable in skills, and what needs a dedicated persona or tool set in an agent - then let subagents handle anything that benefits from parallel or isolated execution.

Summary

GitHub Copilot's customization model is built on three primitives - Instructions, Skills, and Agents - layered on top of a context-window system that you can explicitly shape. Mastering these lets you move from reactive chat to orchestrated, repeatable agentic workflows.

Why it matters

  • Default Copilot behavior is generic; the primitives let you encode your team's conventions, tooling, and workflows so they apply automatically.
  • Understanding context prioritization tells you why suggestions miss and how to fix it fast.
  • Subagent orchestration unlocks parallel or decomposed work that would overwhelm a single chat thread.

How it works in practice

The three primitives (layered)

  • Instructions (*.instructions.md) - persistent background context scoped by glob pattern. Use for coding standards, architecture rules, and compliance guardrails. Applied automatically; no user action required.
  • Skills (folder + SKILL.md) - reusable, bundled capabilities invokable via /command or auto-discovered by agents. Can include templates, scripts, and reference files. Supersede the older *.prompt.md pattern.
  • Agents (*.agent.md) - opinionated personas that combine tools, MCP servers, and instructions. Use for specialist workflows (e.g., "Terraform Expert", "Security Reviewer") or as coordinators that delegate to subagents.

Context — what Copilot sees and how it's prioritized

  1. Code surrounding the cursor
  2. Explicitly #-referenced files (VS Code) or @-mentioned files (CLI)
  3. Recently modified files
  4. Files directly imported by the current file

Copilot does not scan the full repo by default. Use #codebase in chat to force a repo-wide search.

Subagents — when and how

  • A subagent is a temporary worker launched by the main agent for context isolation, specialization, or parallelism.
  • Main agent = project lead (persists, talks to user). Subagent = focused contributor (temporary, reports back).
  • In VS Code: enable the agent tool in frontmatter; use an agents: allowlist to control which workers the coordinator can call.
  • In Copilot CLI: use /fleet to decompose a task into parallel background subagents automatically.
  • Recursive subagent spawning is off by default (chat.subagents.allowInvocationsFromSubagents).

Example

Coordinator + worker agent (VS Code frontmatter)

---
name: Feature Builder
tools: ['agent', 'read', 'search', 'edit']
agents: ['Planner', 'Implementer', 'Reviewer']
---

---
name: Planner
user-invocable: false
tools: ['read', 'search']
---

Scoped instruction file

---
applyTo: "src/**"
---
# Coding standards
- Open braces on the same line.
- No implicit usings.

CLI fleet invocation

/fleet Update auth docs, refactor auth service, and add related tests.

Gotchas

  • Closing a file removes it from Copilot's active context - keep related files open while working.
  • handoffs and argument-hint frontmatter work in VS Code but are ignored on GitHub.com cloud agents - don't assume portability.
  • Skills replace prompt files; avoid investing in new *.prompt.md files.
  • Subagents share the same filesystem - overlapping file writes across parallel agents cause conflicts.
  • user-invocable: false hides an agent from the picker but still allows coordinator delegation; disable-model-invocation: true prevents delegation too.