Don't Panic: Windsurf

Windsurf is an editor where the AI is not parked in a small chat box waiting for a question like an unusually expensive houseplant. Its central character is Cascade, an agent that can read files, propose edits, run commands, and keep track of the work around it. The point is not that code becomes self-writing. The point is that the editor can carry more of the awkward middle: finding relevant files, connecting a request to a change, and checking what happened next.

The important bit is Flow, the session context that follows open files, recent edits, terminal output, Git state, and other clues around the task. That is why a request can start with a question and later become a multi-file change without you pasting the same background into every message. Context is useful, but it is not clairvoyance wearing a lanyard. It describes what the editor has seen, not what the change ought to mean.

Cascade has two working postures. Chat mode is for asking about code and examining an idea without modifying the project. Code mode is for the consequential part: reading, editing, running, and iterating. The distinction matters because the second mode can create a convincing trail of activity. A trail is not proof. Read the proposed patch, inspect the command output, and decide whether the result belongs in the repository.

The durable control is the project instruction. This course calls it .windsurfrules: a file that records conventions such as strict TypeScript, component patterns, test placement, and error-message style. Rules can be global, project-specific, or narrower for a directory. That gives Cascade a map, but a map still needs a human choosing the destination. A beautifully consistent wrong change remains wrong with impeccable formatting.

MCP extends the agent with tools for databases, APIs, and other services. Workflows turn repeated instructions into named routines. Both reduce repeated setup, and both enlarge the consequences of a careless request. Start with a contained task. Ask for an explanation. Move to Code mode when the intended edit and its test are clear. Keep the diff small enough to understand without needing a second agent to explain the first one.

Read the intro for architecture and vocabulary. Use the slides for moving parts and their relationships. Keep the cheatsheet nearby when configuring rules, memory, MCP, workflows, and shortcuts. Then use the quiz to check whether the mental model survived contact with the buttons, which is where most mental models discover they had no shoes.