A loop in your repo · Work with coding agents

Verified 3 Nov 2026claude-code 2.1.317

You've used Claude in a chat. A coding agent is Claude with hands: it works inside your project folder, reads your files, edits them, runs your commands and looks at what happened. Claude Code is Anthropic's coding agent, and it's the one this course and the build-along courses use.

Under the hood it's the same loop you'll build yourself in Your first agent loop: Claude picks a tool, the tool runs, the result goes back to Claude, and it decides what to do next. It keeps going until the job is done or it needs you.

One task, the agent's waystep by step

ReadIt looks around: the files you named, your CLAUDE.md, a search for the code that matters.

01 ReadIt looks around: the files you named, your CLAUDE.md, a search for the code that matters.

02 PlanIt decides what to change. In plan mode it stops here and shows you.

03 EditIt changes files, asking first unless you've allowed it.

04 CheckIt runs your tests, types or linter and reads the output.

05 ReportIt tells you what it did, ideally with the command it ran and the result.

step 1 of 5

To install it on macOS, Linux or WSL (checked November 2026):

Shell

curl -fsSL https://claude.ai/install.sh | bash
cd your-project
claude

On Windows PowerShell it's irm https://claude.ai/install.ps1 | iex. One honest note: Claude Code needs a paid Claude plan or an API (Console) account. The free claude.ai plan doesn't include it. This lesson's exercise runs here for free.

By default Claude Code asks before it edits a file, runs a command or goes online. Press Shift+Tab to cycle modes, including plan mode, where it researches and proposes but doesn't touch your code.

The loop is only as good as your brief. "Fix the login bug" makes the agent guess where to look and when it's done. Give it where to look, what's wrong and how to check it's fixed: a test it can run that passes or fails. Anthropic's best-practice guide says it plainly: if Claude said it was done and wasn't, it usually had no way to check its own work.

Your turn. Moon Café's website shows Sunday as "open" when the café is closed. The opening-hours logic is in src/hours.ts, its tests are in tests/hours.test.ts, and pnpm test runs them. Rewrite the vague brief so an agent could do it right first time, and ask for a plan before any change. This sandbox Claude can't see a repo, so its reply is the plan it would follow.

Your turnRewrite “fix the hours bug” as a brief a coding agent can act on

Goal Brief an agent with where to look, what's wrong, how to prove it's fixed, and a plan first

What good looks likeThe coach checks the reply

Says where to looknot checked yet
Says what's wrongnot checked yet
Says how to check itnot checked yet
Asks for a plan firstnot checked yet
Ran at 1.0not checked yet
Replies with a plannot checked yet
Describes the symptom a user sees, not only “a bug”not checked yet
Gives a pass-or-fail check: a failing test to write, or a command to runnot checked yet
Keeps the job to this one bugnice to have
The reply is a plan of steps, not finished codenice to have

Run your prompt and the replies land here.

Try sending the vague draft first. What does Claude have to guess?

Down the rabbit hole