If the skill is a process then very effective. Do x, then y, then z, then if w do... Whatever.

Or things like style guides, or instructions I find myself typing over and over.

Look at something basic like reading a file that the LLM can't read natively, a PDF or whatever. There's hundred of ways to extract the data but rather than let the agent choose randomly each time I can set up a read PDF skill and specify my preferred way to do it.

You can have LLMs do something via a prompt, then you can try making a skill for it and see how the behavior changes.

IMO still very very useful for repetitive tasks that you want done a certain way.

How would the base model know bespoke organizational policy, process and nuances without some way of briefing it?

Your mileage may vary. So you should experiment.

LLM harnesses haven’t retired the feature