Now you use the loop for real. You will take the app you built in Week 1 and add something to it that a single generation pass could not do cleanly, directed through Claude Code, and you will do it without breaking what already works.

Claude Code is an agentic coding tool: you point it at your project and talk to it in plain language, and it can read across your files, propose a change, and carry it out. The difference from prompting Lovable is one of precision and reach. Lovable is superb at generating and visually iterating; Claude Code is where you go when you want a specific, surgical change ("add a filter to the results list that remembers the user's last choice") and you want to see exactly what it touched.

You will run the loop from the previous lesson live. Describe the one change you want, small and specific. Let Claude Code show its plan, and read it. Let it make the change, then review the diff so you can see precisely what moved. Then, and this is the step that makes it safe, go back to your app and confirm the things that worked last week still work. This is where "verify" becomes concrete: a change that adds your new feature but quietly breaks login is not a win, and the only way you would catch that is by checking. You check.

Expect to redirect it at least once, and treat that as normal rather than failure. The agent will sometimes do slightly too much, or misread what you meant, exactly like a fast junior would. When it does, you do not throw out the approach, you tighten the instruction and go again. Learning to steer a capable-but-literal collaborator is the skill; the feature you ship is the evidence you have it.

By the end you have a real upgrade to your capstone that you directed rather than typed, and, just as importantly, proof that your app still works end to end after the change.