title: "Best Practices" description: "Proven techniques for getting the most out of AI in VS Code" section: "Copilot" readTime: "6 min"

Best Practices for AI in VS Code

Pick the Right Tool

Choose the Right Agent Type

• Local — interactive work needing editor context (test failures, debug output, MCP tools)
• Copilot CLI / Cloud — well-defined tasks you can offload while continuing to work
• Cloud — team collaboration, opening PRs, assigning GitHub issues
• Parallel sessions — run multiple agents simultaneously for independent tasks

Hand off between types: start local to explore/plan, hand off to background/cloud for implementation.

Write Effective Prompts

1. Be specific — state language, frameworks, libraries, expected behavior

Write a TypeScript function that validates email addresses.
Return true for valid, false otherwise. Don't use regex.

2. Include verification — add test cases so the AI can validate its own work
3. Break down complex tasks — one well-scoped step at a time
4. Avoid vague prompts — not "make this better"; say "reduce time complexity" or "add null validation"
5. Iterate — add constraints in follow-up messages instead of rewriting the whole prompt
6. Course-correct early — steer with a follow-up, queue a message, or stop and restart
7. Ask for clarification — instruct the AI to ask questions before proceeding when the task is ambiguous

Provide the Right Context

• Use #file, #folder, #symbol to reference specific code
• Use #fetch for up-to-date information from web pages
• Reference source control changes (#changes), terminal output, or test failures
• Attach images or screenshots for visual context
• Open related files so Copilot has broader codebase context

Choose the Right Model

• Fast models — simple completions, boilerplate
• Reasoning models — planning, debugging, architecture decisions
• Pin a model in custom agent or prompt file definitions for consistency
• Experiment: different models produce different results for the same prompt
• Use BYOK (Bring Your Own Key) for additional model options

Plan First, Then Implement

For complex multi-file changes:

1. Explore — use Ask mode to understand the existing code
2. Plan — use the Plan agent to create a structured implementation plan
3. Implement — switch to Agent mode, include tests or expected outputs
4. Review — use checkpoints to roll back if the agent goes off track

Review AI Output

• Read generated code before accepting — check edge cases and error handling
• Run tests after every AI change
• Use checkpoints (chat.checkpoints.enabled) to roll back to before a change
• Check for security issues: injection flaws, hardcoded secrets, missing input validation
• Never paste credentials or sensitive data into prompts

Manage Context and Sessions

• Start new sessions for unrelated tasks — context pollution degrades quality
• Delete irrelevant chat history or use /compact to summarize it
• Use subagents for isolated research that shouldn't pollute main context
• Run parallel sessions for independent tasks via the Sessions list

Large Codebases

• VS Code auto-indexes with semantic search + language intelligence
• Enable remote indexing for large GitHub repos
• Use multi-root workspaces to scope context in monorepos
• Add custom instructions describing architecture and module boundaries
• Use the Plan agent for cross-cutting changes spanning many files