The Twenty Percent Problem

One of the biggest struggles I've had with AI projects is what I call the twenty percent problem — Pareto's principle in action. AI can get you to eighty percent almost instantly. With just a little effort, you can generate ideas, boilerplate, basic implementations super quickly. But that remaining twenty percent? The refinement, the small decisions, the details that actually make something feel intentional and polished — that hasn't gotten any faster.

Ideas don't come fully baked. When you ask an agent to generate a project concept, it can do a solid job with the generic version. But you haven't actually thought through all the small details yet. And those details require reps, iterations, constant tiny decision-making.

I'm not sure agents can do that for you unless you're comfortable just shipping something okay. But if you want to build good product, you have to own that twenty percent. Maybe there are specialist tools — compound engineering approaches, for instance — that can help you review and evaluate what you've built. But you still need a human in the loop. You have to step through all those decisions yourself. And that takes time.

The twenty percent problem is real, and it's where most of the actual work lives now. That's been the thing I've struggled with the most.