After playing around with a bunch of fancy skills recently, the one with the most compound value is maintaining an experience skill or an MD file. It requires the agent to automatically review and consolidate lessons learned after each task, so that historical experience is reviewed first when executing a new task next time.

Now, using the same model, I have it record what I care about each time I communicate my needs to it, maintaining my communication preferences. It seems communication efficiency has improved quite a bit.

Project experience and personal preferences are all context. Mining, storing, and utilizing context is the most important thing.