Reading7 min read·Module 1

One sentence in, a verified fix out

Before any definitions, watch one real session. A login test has been failing since yesterday. The person types one sentence and then does nothing else. Step through it with Next step, and keep an eye on the counter in the corner that tracks how many times you had to act:

one bug, two ways

claude — ~/shop-apiyou acted 1×

> The login test is failing since yesterday. Find out why and fix it.

▮ …

You acted once; Claude did the rest and proved it. Every line after your prompt is Claude acting through a tool, labeled with its loop phase. Notice the red failing run: the agent reads its own failure and keeps going, and the loop only ends on a verified pass.

Everything after the prompt was Claude working through tools in a real project: it searched for the test, read it, ran it, read the failure, found the handler, changed one line, and ran the test again to prove the change worked. On the Chat tab the model is exactly as smart, and you still do six round trips by hand, because a chatbot hands you advice while an agent does the work and checks that it worked.

That is the whole difference. Claude Code is an agent: a model connected to tools, running in your terminal, that can read your files, search your codebase, edit code, and run commands itself. You describe the outcome you want, and it does the work where the work lives.

1 / 4