Why is it so hard to get an AI-built skill or workflow fully working, and what should I do differently?

The group's take was that this mirrors normal software development — prototyping is easy, but making something fully reliable is the real work, and professional developers would likely say 'welcome to what we do.' Suggested fixes discussed: scale back scope so the AI does less (e.g., have it complete only part of a document perfectly rather than the whole engagement letter), and take one problem at a time — perfect it before moving to the next, potentially scaling by assigning different team members to different problems. It was also noted that there are many ways to approach a build, and sometimes you hit a dead end and need to pivot to a different approach rather than pushing forward on the original path. Simplifying the overall approach was seen as generally preferable.