How do you guide it when you have no idea yourself? You mean you must first reverse engineer it manually and then hint it to the solution? Why would you do that at that point?

I feel like you might still be perceiving the AI suggestion as some magic box that just does things. It's not. It's a very fast very eager junior.

Why and how do people hand work to juniors?

You have a rough idea where you want to go, you have a rough idea what needs to be done, and then you iterate.

For example, you think "okay, I need some way of validating this" and then you tell it to run the software to generate test data.

While errors may creep in, you should be able to validate that step.

And then you use that test data to validate whatever next steps it should be doing.

It's essentially the same workflow as any mapping out of something new, but with the individual mapping steps being done by a drone. You do still draw the actual map, do logistics and strategize.

It's not easy work, but it's also not difficult work. What it is is laborious, and that's where LLMs help.

__

To stick with your database format question:

If the software uses that data, it also needs code to parse and interact (with) it.

This code resides within it, and you can forcefully pull it out.

That step of "write me a spec for this data format based on this code. Build a parser in python and build test cases using this DB file" can be done agentically (when sliced into small logical blocks of work orders).

You see how you went from "it's easier to use ai instead of using wine on linux" to "to use ai you must do all this work"

Well then using wine was the correct solution to begin with?

> This code resides within it, and you can forcefully pull it out.

What code? If you had the code you could port it.