You have an idea for a small App. An AI builder can turn a few sentences into a polished screen surprisingly quickly—but a screen that looks finished may still behave differently from the brief.

This Course is a guided audit of one end-to-end, low-risk project: a mobile-first checklist for preparing and closing a community book-swap event. The Course supplies twelve synthetic tasks containing no real personal or organization data. Before each production result is revealed, you make the relevant scope, testing, repair, approval, or handoff decision and compare it with evidence from the real run.

You will first compare the main AI coding work modes and representative products, then follow one continuous Replit-to-Codex path. You begin in Replit because there is no existing code and the first need is a runnable user path with little setup. You will turn requirements into observable checks, classify the results, and use a real failure to form one bounded repair request instead of repeatedly asking AI to “make it better.”

Later, you inspect a file-level conflict that the visible Preview could not settle. That concrete evidence—not the idea that a Coding Agent is always better—determines whether a move to Codex is justified. You preserve the working state, ask Codex to inspect before editing, approve an exact scope, check the actual difference, and interpret both successful and blocked checks honestly.

The learning investigation ends with a reviewable, non-public demonstration and a clear handoff, even though the original App brief is not fully satisfied. The Course does not claim that the App is production-ready, independently reproducible from an incomplete export, or proven safe for real data. You leave with a transferable method for choosing an AI coding tool, limiting a first version, testing behavior, requesting focused changes, checking Agent work, and stopping where the evidence stops.