Prompt-to-Output Ratio

At some point while working with AI, I noticed a pattern I now take fairly seriously. I call it my prompt-to-output ratio. Very simply: How much input do I put in—and how much output comes out? That sounds technical at first, but for me it has become a pretty good signal for whether I am producing something useful or just churning out AI slop.

How I recognize good results

I recently created a psychology article for my site with AI. The whole process, from first idea to published version, took maybe 15 to 20 minutes—including German and English versions. That is absurdly fast.

What mattered was not speed but quality. When I read the text through, it was clear: This is usable. It has substance. It is not generic AI mush.

Then something struck me: The prompt I wrote was longer than the finished article.

When the prompt is longer than the output

That was not a one-off. I see it again and again, including when programming. I write long prompts, sometimes several in a row, describing the architecture, the requirements, the background thinking. And in the end maybe ten lines of code come out—but exactly the right ones.

When I look at the total amount of input I provided, it is often larger than the output the AI produced. And in those cases, the result is almost always good.

Because I already did the thinking.

When I think too little, slop comes out

The flip side is just as clear. When I only write: "Write me an article about X" or "Build me a UI for this data," I get output that sort of fits—but is also completely interchangeable.

That is what people call AI slop.

And it is no accident. It happens because I was sloppy myself. I made no real decisions, offered no perspective, set no constraints. I asked the AI to guess what I might mean.

And that is exactly what I get: a statistical average.

What I am actually measuring

It is clear to me now that with this ratio I am not really measuring length, but something else: How much real thinking comes from me?

A long prompt is no guarantee of quality. But in practice it is often a good sign, because many decisions are packed into it. I have thought through what I want and what I do not, who it is for, what the purpose is.

The AI no longer has to "think"—only structure, shorten, and phrase.

A simple red flag

From that, a pretty practical heuristic emerged for me.

When the prompt-to-output ratio is clearly negative—that is, when I give very little input and demand a lot of output—that is a red flag. Not proof the result is bad, but a strong warning signal.

Then I should ask: Have I actually thought this through enough?

A green flag, not proof

The reverse is a green flag for me. When my input is more extensive than the output, the odds are high that I already did the substantive work and the AI is helping me shape it cleanly.

That is no guarantee of quality. I can pack a lot of nonsense into a long prompt. But in practice it correlates surprisingly strongly with good results.

AI as editor, not thinker

At core, for me it comes down to a question of roles.

When I give little input and demand a lot of output, I treat the AI like a thinker. I delegate the actual work.

When I give a lot of input and have the AI compress, structure, and phrase, I use it as a tool. As an editor. As a kind of compiler for my thoughts.

That is when it becomes truly strong.

Conclusion

For me, the prompt-to-output ratio is not an exact measure, but a simple diagnostic tool. A kind of self-check.

When I notice I want a lot of output from little input, I know: I am trying to skip work I should really do myself.

And when I see that my input is substantial and considered and the output becomes compact and precise, that is a good sign.

In the end the rule is simple: If I want good results, I have to think first myself. The AI can speed that up—but it cannot do it for me.