Claude Code Tutorial: From Install to Your First Shipped Feature

Set up a workstation the agent can succeed in

An agent is only as good as the environment it works in. Before your first prompt, make sure the repo has a working test command, a formatter, and a clean Git history so every agent change is reviewable as a diff.

• VS Code with the terminal you actually use.
• Git initialised, with a branch per agent task.
• Node.js or Python installed and on PATH.
• One command that runs the test suite.

The prompt pattern that works

Give context, constraints, and a definition of done. 'Fix the login bug' fails. 'Users with an expired session see a blank page on /dashboard. Reproduce with the failing test in auth.test.ts, fix the redirect, do not touch the router config' succeeds.

State what the agent must not change. Negative constraints prevent most runaway edits.

Plan, act, verify

Ask for a plan first and read it. Cheap to correct at that stage, expensive after twenty file edits. Then let the agent act, and always require verification — a test run, a type check, a build — as part of the same task.

The verification step is what turns an AI toy into an AI colleague.

Common Claude Code mistakes

These four account for most bad outcomes we see from new users.

• Prompting for a whole feature at once instead of one reviewable slice.
• Skipping the plan and going straight to edits.
• No tests, so nothing can verify the change.
• Accepting a diff you have not read.