VS Code AI: Choose an Assistant and Set Up Project Context

VS Code has a growing set of AI assistants available as extensions: GitHub Copilot is built into the editor's own AI features, Cursor is a VS Code fork with deeper repository indexing, and Claude Code runs in the terminal with a VS Code companion. Each has a different approach to how much of your project it reads and how much it can change at once.

Tools like fenbs keep your project requirements and handoff context outside any one AI assistant, so switching from one to another does not mean starting the project explanation from scratch.

Which AI assistant works in VS Code

GitHub Copilot is the most tightly integrated option: it is available directly in VS Code's settings and has a chat panel, inline completions, and an agent mode for multi-file edits. Extensions for other assistants are available through the marketplace, including a Claude Code companion extension.

Cursor is a separate editor based on VS Code. It is not a VS Code extension; you install it as a standalone application and it shares most of the VS Code extension ecosystem. It adds full repository indexing and its own agent mode on top of the VS Code foundation.

Setting up project instructions in VS Code

GitHub Copilot reads a file at .github/copilot-instructions.md in your repository root. Put standing rules here: libraries your project uses, patterns to follow, what a reviewer needs to see before a change ships. Keep it short. A list of ten specific rules is more reliably applied than three pages of background.

Claude Code reads an AGENTS.md file at the repository root and in subdirectories. If you use both assistants, you can maintain both files or write shared content that applies to either. Check each assistant's documentation for the current file name and format.

What to put in project instructions

• The project's technology choices and why: "We use TypeScript with strict mode; do not add any-typed variables."
• Patterns the AI should not break: "All state changes go through the service layer; do not write directly to the database from a component."
• What the reviewer needs to see: "Every PR must have at least one test for the changed function."
• What the AI should not attempt: "Do not modify the migration files; hand those back for a person to write."

Using agent mode safely

Agent mode in GitHub Copilot and Cursor can edit multiple files, run builds and apply a complete diff. It is useful for well-scoped tasks where you have written clear requirements. Before running agent mode, write the task in one sentence. After it runs, check the diff against that sentence. If it touched files outside the stated scope, ask it to revert those changes before you commit. A well-reviewed diff takes a few minutes; an unreviewed diff can introduce a subtle bug that takes hours to find.

Agent mode produces a large diff quickly. A quick diff review is still a diff review: check for changes you did not ask for, and ask the assistant to explain any change you do not understand.

Keeping context when you change assistants

Chat history stored inside GitHub Copilot stays in GitHub Copilot. When you switch to Cursor or Claude Code, that history is not available. Requirements and decisions that live only in a chat panel do not travel.

Record the decisions that matter in a place every assistant can reach. What is an AI brain explains why context outside a single AI app is more reliable than depending on any assistant's built-in memory. A team that keeps requirements and rules in fenbs can onboard a new developer on a new machine in minutes: connect the AI, read the project context, start working.

Reading context before each session

One habit that improves AI-assisted coding sessions significantly is asking the assistant to read the project context before it starts. This means the session begins with the agreed decisions in scope rather than the assistant inferring them from the code alone.

In practice: start each session with a short message like "read the project context, including rules and where we left off, then tell me what the next task is." If the assistant's answer matches the record, proceed. If it does not, the context may be outdated and worth updating before the session continues. This habit catches inconsistencies before they turn into wrong code rather than after. It also makes sessions shorter: the assistant starts with shared understanding rather than building it incrementally through the first several exchanges.

How to do this with fenbs

1. Create a free account and add your current project.
2. Write the key requirements and decisions in fenbs: what the project is for, what the team has agreed not to do and the current tasks.
3. Connect your VS Code AI assistant over MCP from Integrations. GitHub Copilot, Claude Code and Cursor are supported.
4. Ask the assistant to read the project context at the start of each session before it edits code.
5. Before you switch assistants, ask the current one to save progress: where each task stands and the next step.