The seemingly simple act of asking an AI a question can be systematically managed. Incorporating AI prompting into the PDCA cycle turns random, hit-or-miss queries into a standardised, repeatable, and controllable closed-loop process.
The Plan phase is about defining the problem boundary. Before asking the AI anything, clarify what specific problem you are trying to solve, what format you expect the output to take, what information isknown, what is unknown, and how you will judge whether the output is useful. A common mistake is describing the problem too vaguely, causing the AI to produce lengthy output in the wrong direction. Before the AI receives any input, the human must first be clear about what they want.
The Do phase is executing the query. Based on the plan, issue a structured instruction to the AI. The structure of the instruction determines the quality of the output. Provide context, specify constraints, define the format, give examples. These are all effective ways to improve output quality. The AI is not a mind reader. The quality of its output depends heavily on the quality of its input.
The Check phase is evaluating the output. AI output is not necessarily correct, and it does not necessarily meet your expectations. Evaluate it from three dimensions: accuracy, completeness, and practicality. For important questions, cross-validate the AI conclusions. What AI is good at is generating plausible-looking content, not guaranteeing truthful content.
The Act phase is taking action based on the evaluation. If the output passes, move to the next stage. If it has flaws, refine the input and query again. If the direction is completely wrong, go back to the Plan phase and redefine the problem.
The core value of this cycle is that it turns every AI interaction into a learning opportunity. After each cycle, you understand a little more about how to ask better questions. The more you use it, the more precise it becomes. The more you use it, the less effort it takes. This is the optimal state of human-AI collaboration.
In practice, the principle can be summarised in sixteen characters: humans set the rules, AI executes, humans verify, iterate to optimise.