I wonder if generated skills could be useful to codify the outcome of long sessions where the agent has tried a bunch of things and then finally settled on a solution based on a mixture of test failures and user feedback