I think skills actually complement what you're saying very well.

> You get way more interesting behavior from agents when you allow them to probe their environment for a few turns and feed them errors about how their actions are inappropriate. It doesn't take very long for the model to "lock on" to the expected behavior if you are detailed in your tool feedback. I can get high quality outcomes using blank system prompts with good tool feedback.

My primary way of developing skills (and previously cursor rules) is to start blank, let the LLM explore, and correct it as we go until the problem is solved. I then ask it to generate a skill (or rule) that explains the process in a way that it could refer to to repeat this again. Next time something like that comes up, we use the skill. If any correction is needed, I tell it to update the skill.

That way we get to have it explore and get more context initially, and then essentially "cache" that summarized context on the process for another time.

Error feedback from tools could be argued to be isomorphic with skills (or the development of them). It tracks with how we learn things in meatspace. Whatever strings we return in response to a bad SQL query or compiler error could also include the contents of some skill.md file.