The quality of AI-assisted software depends on unit of work management

The author argues that the quality of AI-assisted software depends mainly on how well units of work are managed, since context and compounding errors matter more than raw model intelligence. Using a 5% per-action error rate, they calculate a 59.9% chance of success over 10 turns, and they argue that METR's reported success rates fall to around 40% for 2-hour tasks once real-world messiness is considered, with GPT-5 cited as the top model. They propose user stories as small units of work that deliver legible business value at each checkpoint, and describe an experiment called StoryMachine that currently reads PRDs and Tech Specs to produce story cards.