I told one AI to demolish the handoff prompt I wrote for another AI. It found a test that passes even when it's empty

A reporter, while working on a project with Claude Code, faced an unsettling issue. The commander session wrote the instructions, while separate worker sessions implemented them in parallel. The commander's task was to create a handoff prompt that the worker could read and execute. The handoff prompt was deemed the most dangerous part of the process, as the worker trusted it as the specification and began working immediately.

Before handing over the instructions, the reporter ran them past a subagent designed to find flaws and not approve. This time, however, the demolition agent discovered something alarming. The first objection was that the test defined as "Get the XX test passing (green)" was flawed. The demolition agent found that the test only checked if the process ended, not if it succeeded.

The second objection revealed that the config flag to switch in the real component did not exist, and the function it referenced was never called. The reporter, confident in their instructions, had not double-checked the code before handing it over. This incident highlights the importance of thoroughly reviewing the code before delegating tasks, as relying solely on written instructions can lead to significant problems.