[satisfice]

The skills list weirdly leaves out “knowing the capabilities and limitations of the tools.” This is non-trivial to say the least.

The roles list also leaves out testing, which seems to me to be the second most important thing (after specifying). This may be because non-testers assume that testing is easy or will be done by the AI. But any testing done by AI is not testing at all, because the effect of real testing is to inform a human of the status of the product based on that human’s empirical investigation. When AI “tests” the humans are being asked to trust instead of investigate.