GitHub Copilot

Copilot vs Copilot Chat vs Copilot Workspace

GitHub Copilot has evolved into a family of products:

Copilot (inline suggestions): The original product. As you type, Copilot suggests code completions ranging from a single token to entire functions. Triggered by natural language comments, function signatures, and existing code patterns. Available in VS Code, JetBrains, Neovim, and Visual Studio.

Copilot Chat: A conversational interface embedded in your IDE. Ask questions about code, request explanations, generate tests, debug errors. Chat has access to the current file and selected code. Available in VS Code and JetBrains.

Copilot Workspace: GitHub's agentic environment — enter an issue or task description and Copilot proposes a plan, edits files, and runs tests to solve it. Competes with Cursor's Agent mode. Still in preview as of 2024.

How Copilot completions work

Copilot constructs a prompt from your current file context and sends it to the LLM (GPT-4o or Codex variant). The prompt includes: the current file content (truncated to fit context), a "cursor position" indicator, open tabs in the editor (nearby files), and your comment/function signature.

# Example: how Copilot uses your comment as a prompt
# Function to parse a date string in multiple formats
# Supports: YYYY-MM-DD, MM/DD/YYYY, DD Mon YYYY
def parse_date(date_str: str) -> datetime:
    # Copilot generates the implementation based on the comment above
    from datetime import datetime
    formats = ["%Y-%m-%d", "%m/%d/%Y", "%d %b %Y"]
    for fmt in formats:
        try:
            return datetime.strptime(date_str, fmt)
        except ValueError:
            continue
    raise ValueError(f"Unrecognised date format: {date_str}")

Write descriptive comments, clear function signatures, and use good variable names to get better suggestions. Copilot is essentially doing few-shot learning from your codebase style.

Prompt engineering for Copilot

Copilot responds to context signals. Techniques that improve suggestion quality:

• Write the function signature first: A precise type-annotated signature (especially in Python/TypeScript) constrains the solution space.
• Add a docstring before implementing: Write the docstring (what the function does, parameters, return value) before the body. Copilot reads the docstring as a specification.
• Open related files in tabs: Copilot includes recently opened files in its context. If you're implementing a method, have the model class and related utility files open.
• Use examples in comments: # Example: parse_date("2024-01-15") → datetime(2024, 1, 15) guides the implementation.
• Chain completions: Accept a completion, then start typing the next piece. Copilot learns from what you accepted.

Copilot for the CLI

# Install GitHub CLI and Copilot extension
gh extension install github/gh-copilot

# Ask questions in natural language
gh copilot suggest "list all docker containers using more than 1GB memory"
# → docker stats --no-stream --format "table {{.Name}}	{{.MemUsage}}" | awk '...'

# Explain a command
gh copilot explain "find . -name '*.py' -exec grep -l 'import torch' {} +"
# → "This command finds all Python files that import PyTorch..."

# Interactive mode
gh copilot suggest -t shell "compress all jpg files in current directory"

Copilot CLI is useful for remembering complex shell syntax, constructing awk/sed/jq pipelines, and crafting git commands for unusual operations.

GitHub Copilot Enterprise features

Copilot Enterprise (available on GitHub Enterprise Cloud) adds organisation-specific features:

• Organisation-wide knowledge base: Index private repos, wikis, and documentation. Copilot Chat answers questions against your company's internal knowledge ("How do we handle authentication in our API?").
• Fine-tuned models: Train Copilot on your private codebase to suggest code in your organisation's style and use your internal libraries.
• Code review: Copilot reviews pull requests and leaves comments on bugs, style issues, and security vulnerabilities.
• Data privacy guarantees: Enterprise tier ensures code is not used for model training and offers a Business Associate Agreement for regulated industries.

Privacy and security

Copilot's data handling is a common concern for teams using it with proprietary code:

• Copilot Individual: Code snippets may be sent to Microsoft for model improvement (opt-out available in settings).
• Copilot Business: Code is not used for training. Prompts and suggestions are encrypted in transit and not stored beyond the request.
• Copilot Enterprise: All Business protections plus additional audit logging and enterprise compliance features.

Security concern: Copilot can suggest code with vulnerabilities if the training data contained vulnerable code. GitHub added a "public code filter" that blocks suggestions that closely match public code (to avoid licence issues), but this doesn't prevent novel vulnerable code. Always code review AI-generated code, especially for security-sensitive paths.

Gotchas

Hallucinated libraries: Copilot occasionally suggests imports of libraries that don't exist or function calls with wrong signatures. Always verify generated code runs without import errors before committing.

Context blindness: Copilot's context window for completions is limited (~8K tokens). For large files, it only sees the nearest surrounding code. In a 5,000-line file, Copilot completing code in the middle won't see definitions at the top of the file. Break large files into modules.

Telemetry in free tier: The free Copilot tier (introduced 2024) includes 2,000 completions/month but may have different privacy terms than Business. Read the ToS before using with sensitive code.

Not a replacement for tests: Copilot-generated code can look correct and fail at edge cases. Always write tests for generated logic, especially parsing code, string manipulation, and boundary conditions.