A new AI conversation window starts with some initial state assumptions around response formatting, engagement tone, continuity, etc. But essentially, all permissible knowledge is in play. The trick isn’t getting AI to be smarter but to constrict the information pool so responses are more accurate and relevant to your task at hand.

If, for instance, you are planning a large party and use one window or state to get all of the arrangements mapped out, you will get better, faster, and more accurate responses if you feed just the menu into a new window to generate the recipes. This simplifies the state, removes irrelevant possibilities, and provides a more focused picture of the task at hand. Low quality results from AI often correlate with how much latitude the system is given.

Unlike a person who would ask clarifying questions to the above request, the AI may not. It will most often lean toward answering, often with a large number of possibilities. While it may sound counterintuitive, the more constrained and defined a state is, the better the output for the current task.

The key isn’t maximizing the instructions, it is minimizing the irrelevant possibilities.

In this series, I will touch on the many ways to influence and control state as well as its benefits. While some are harder than others, there is a simple and effective technique that is useful in conjunction with all the other techniques. Use multiple windows sequentially that have increasingly constrictive states.

When I develop, I start with a window to explore ideas and give form to my concept. I work in the window to create a project plan and a prompt for the next window. The new window uses these two elements for its starting state and builds a technical spec and several prompts for the next group of windows to write the final, detailed specs. I then take these documents, put them into an isolated build window, and create a state with minimal ambiguity and reduce the need for AI interpretation.* This method has resulted in incredible accuracy, consistency and autonomy.

It is important to note that a new window clears the accumulated conversation, but depending on the system, persistent instructions, memory or other state may still be carried into it.

The key isn’t maximizing the instructions, it is minimizing the irrelevant possibilities. The following articles look at some of the most effective methods for establishing and maintaining control of your states.

*I also instruct the AI to never guess. “When uncertain, don’t guess; ask for clarification.”