03 / 10
Who checks the work?
npx halfcycle Not the thing that did it. Whatever wrote your plan will tell you the plan is fine, because it is the worst available judge of its own output. Halfcycle reads the work back with something that did not write it, against conditions agreed before the work started, and you get a record of what was checked.
Nobody grades their own homework.
For most people building with AI tools today, the honest answer is the same tool that did the work, or nobody.
That is not laziness. It is the obvious thing to do. You ask the tool to write a plan, then you ask it whether the plan is good, and it says yes. The trouble is that it wrote the plan from its own understanding of your problem, and it reviews the plan against that same understanding. Any gap in how it understood you is in both places, so it cannot see it. People work the same way, which is why editors exist.
A small example. You ask for a booking system for a clinic. The plan assumes every appointment lasts thirty minutes. When you ask the tool to review its plan, it confirms the thirty-minute slots add up and the screens are consistent. They are. It does not ask whether some appointments take an hour, because it never thought they might.
Halfcycle has the work read back by something that did not write it, against conditions you agreed before the work began, such as appointments being allowed to run different lengths. You get a written record of what was looked at and what was found, so you are not relying on anybody’s word that it was fine.