>> prompts are not loops
The prompt has become the visible unit of AI work. We polish the wording, add roles, list constraints, and search for the sentence that will make the model behave. This can improve one response. It does not make a system.
A prompt describes an attempt. A loop gives that attempt somewhere to go when reality disagrees. It gathers context, acts, reads the result, checks the result against a condition, and either stops or tries again with new evidence. The intelligence is not only in the request. It is in the return path.
This changes what good instructions look like. The model does not need one enormous prompt containing every possible decision. It needs a clear next action, access to the relevant state, and a way to discover whether the action worked. A compiler error is better than another paragraph of encouragement. A failing test is context with teeth.
The loop also needs restraint. Without a budget, a stop condition, and a definition of success, iteration becomes motion without progress. The model can keep rewriting the same answer because nothing outside the conversation is allowed to settle the question.
Prompts still matter. They set intent, boundaries, and tone. But reliability comes from the surrounding machinery: inspect before acting, make one bounded change, run the check, preserve the evidence, and return the failure to the next attempt. The prompt starts the work. The loop makes it accountable.
The useful question is therefore not “what is the perfect prompt?” It is “what will the system learn after the first answer is wrong?”