Restarting Rather Than Continuing

Artificial intelligence that writes text is called a Large Language Model. In English, it’s written as LLM for short. OpenAI’s GPT, Google’s Gemini, and Anthropic’s Claude fall into this category. Hereafter, I will call them “models.”

Models do not remember previous conversations. The reason it looks like a conversation is continuing is that everything up to that point is being resent every single time.

From this, one actually effective technique emerges: it is better to cut the thread and start over than to maintain a long conversation.

Continuing makes all three things worse

First, it gets slower. By the tenth exchange, you are sending nine full exchanges. Even if the single response coming back is the same length, the amount of text it has to read to get there is different.

Second, it gets more expensive. Fees are determined by the amount sent. Even for the same question, asking it in the latter half of a conversation is more expensive than asking it at the beginning. The feeling that your credits suddenly start disappearing faster in the latter half of a long conversation is not an illusion.

Third, accuracy drops. The “window” is filled with a mountain of exchanges from the first half, most of which have nothing to do with the current task. Just like when asking a person, attaching a massive amount of irrelevant material can actually cause them to miss the mark. Furthermore, if you changed your direction halfway through, the conversation from before the change remains in the window. You are asking for an answer while both the old policy and the new policy are present.

What to take with you

When starting over, there is no point in taking everything as-is. Take only what you need.

What you take usually consists of these three things:

Decided matters. What are you making? What did you decide not to do? These are decided at the start and continue to be effective throughout.

Current state. How far have you progressed? The names and locations of the things you have on hand.

The next step. A single line is enough.

Conversely, it’s clear what you don’t need to take: the trial and error along the way, ideas that didn’t work, and the times you rephrased things. Those are processes; they are already subsumed in the conclusion.

When you write it out, it usually fits on a single sheet of paper. It might seem like over-compressing to turn two hours of conversation into one page, but that is all that is actually necessary.

The act of writing it out is effective in itself

And there is a byproduct to the task of writing this down.

In the middle of a long conversation, you are surprisingly unaware of what you have actually decided. It is only when you try to write it out that you notice the items that remain undecided. When you find yourself unable to write the page to take with you, it’s usually because the conversation itself has become lost. You end up sorting that out before restarting.

While you are continuing, that state of being lost is invisible. Because the previous exchanges remain in the window, the conversation seems to proceed plausibly.

We create the boundaries

The model will never tell you, “Let’s wrap this up soon.” It cannot see from its side that the window is filling up or that your direction has changed halfway through.

Therefore, creating the boundary is a human’s job. As a rule of thumb: when the topic changes, when you change direction, or when one task is finished. Each time, write a single page and start fresh.

It may seem tedious, but compared to the price you pay for continuing (slowness, cost, inaccuracies), it is usually a bargain.