Software Engineer Shares How Vague Task Prompts Cause AI Execution Failures

Anton, a software engineer working with PHP/Symfony and Go, has documented how imprecise language in task prompts leads to predictable failures when executed by AI models or other developers. He found that phrases like 'take the contract from the neighbouring spec' or 'agree on the approach' function as unresolved decisions rather than clear instructions, forcing the executor to make choices without the author's context. In one example, three such phrases in a single task caused the executor to read wrong files, pick an unintended sample, and ultimately return a clarifying question having produced nothing. His proposed fix is a strictly structured prompt format: a closed file list, a maximum of three explicitly named readable files, no repository searching, and a single named acceptance command scoped to one package. He argues that the failure in such cases lies not with the executor but with the ambiguous sentence that permitted a decision to be made independently.