Claude Code Setup 101

A minimal setup to use Claude (or similar models) effectively for coding in 2025-2026.

1. Choose your interface.
   Claude in the browser (claude.ai) is enough to paste code, ask for changes, and iterate. For deeper work, use Cursor, Windsurf, or Claude Code (when available), they give the model access to your project, terminal, and files.

2. Give it structure.
   In the project root, add a short AGENTS.md or CLAUDE.md: stack, main folders, key files, and a few rules (e.g. "edit existing files when possible," "check package.json for versions"). See Starter Template: agents.md.

3. Scope the ask.
   One clear task per turn works best: "Add error handling to this function" or "Convert this component to use the new API." Avoid "refactor the whole app" in one go.

4. Point to current docs.
   For frameworks and APIs, say "use the official docs for [X] as of 2025" or link the doc. Don't rely on training cut-off dates.

5. Verify and iterate.
   Run the code, run tests, and if something's wrong, paste the error and the relevant snippet. Let the model fix in small steps.

6. Keep a devlog.
   When you or the model fix something non-obvious, note it (what, why, how). Next time you can say "check the devlog for the auth fix" and keep context.

No single "right" setup, this is a baseline. Adjust for your stack and style. For more patterns, see 9 Things for Agent-Driven Development.