You install Zod and write a schema: a description of the data that still exists while the program runs. You infer a type from it, so the type and the check never disagree, and use it in the todo loader. You see Zod refuse a todo with an empty title before the program prints anything, read errors that name the row and the field, and use safeParse to skip a bad row and keep going. Then you write the same schema in Valibot, look at the Standard Schema interface both libraries share, and write one helper that accepts either.

You fetch JSON with Node's built-in fetch from a server on your own machine, and see that its body is unknown. You see that a 404 is still an answer, check the response status before reading the body, and test it with Node's test runner. You see that any value can be thrown, so a caught error is unknown, and narrow it before using it. Then you read environment variables that may be undefined, give them defaults and checks, and validate the todo loader's settings with Zod.

You add the code that works on checked data: a counter of open todos and a list of todos with their status, with no checks of their own. You write tests for valid data, for invalid data, and with fixtures kept in files that go through the real loader. Then you read the project's brief, see the finished program count a valid file, refuse an invalid one with one clear message and exit code 1, and treat a missing file as no todos, proven by one check with five steps.