Write prompts that get the answer right the first time. Every lesson is built on a published guide (Anthropic, OpenAI, Claude Code) and ends with exercises: pick the stronger prompt, or fix a weak one until it passes the checker.
There is no AI behind the exercises: the checker reads your prompt and looks for the parts a strong prompt has. Where a model answers, its reply is pre-written for the part your prompt is missing, so you can see what that part changes. It is not a live run.