For a couple of years, getting good output from a language model required knowing specific tricks – phrasings that worked, structures that helped, incantations passed around in forums. Job listings appeared for it.
That skill is depreciating quickly, and for a good reason.
Why it existed
Early models were sensitive to phrasing in ways that had nothing to do with the task. Asking the same question two ways produced different quality, and nobody could explain why.
People developed workarounds empirically. Some of the tricks were real, most were superstition, and telling them apart required testing that almost nobody did.
Why it is fading
Models got more robust to phrasing. The gap between a carefully constructed prompt and a plainly written request has narrowed to the point where the effort no longer pays.
Tooling absorbed the rest. Applications that use models now handle the structuring internally, so users write normally and the system does the formatting.
What replaced it
Evaluation. The valuable skill in production is not writing the prompt, it is knowing whether the output is good – building test sets, measuring against them, and catching regressions when a model updates.
That is a much less glamorous job and considerably more durable, because it is really quality engineering applied to a probabilistic system.
What still matters
Being specific about what you want, giving examples of good output, and stating constraints. That is not a trick, it is clear communication, and it works on people too.
The general lesson
Skills that exist because a tool is awkward disappear when the tool improves. That is the intended outcome, and building a career on one is a bet against the tool getting better.