Messy and crisp

A colleague reformulated what I do. I had been talking about voice as raw material, about how everything I say ends up as text, about how information becomes denser when it isn't filtered through a form or a brief. He listened for a while and said: "Kladdigt och krispigt." Messy and crisp.

I stopped. Because that was it. Four words that replaced a paragraph of explanation.

Then he added — almost in passing — that it was really "knisprigt." Not krispigt. Knisprigt. A word he had invented in the moment, a portmanteau that contained both modes at once: the organic and the precise, fused together. I asked what he meant. He said that crisp implies cleanliness, clarity, a final state. Knisprigt is in motion. It snaps into place. It is the moment when unstructured material suddenly has form — not because structure was imposed, but because it emerged from the material.

I had never had a word for that. Now I have three.

What it is actually about

The interesting thing is that the distinction is not primarily about what comes out of AI. It is about what you put in.

There are two modes for voice input, and they are not interchangeable.

Messy input is when you have no strict goal. You describe freely. You associate, digress, return, lose the thread and find it again. You let AI interpret — semantic interpretation, assumptions, inference. You say "we want to understand where work gets stuck in the flow" without defining what stuck means or which flow you mean. That is not laziness. It is a deliberate strategy. Messy input is powerful when you are exploring, when you are describing a problem you do not yet fully understand, when what you want is for AI to reflect and structure what you already know but have not formulated.

Crisp — or knisprigt — input is something else. It is when you know exactly what you want. "Add a field for issue type." "Remove this section." "Move the priority to the leftmost column." No assumptions are welcome. No interpretive latitude is requested. You are surgical. And AI is expected to be too.

The mistake most people make is confusing the modes. They give crisp instructions when they are actually exploring — and produce output that is precisely wrong in a confident way. Or they give messy input when they need a specific result — and wonder why AI "doesn't understand what they mean."

Knisprigt is the word for the moment when you deliberately move from one mode to the other. Not a final product. A transition.

Five minutes before, five minutes after

In practice, this is what it looks like for me.

Before a meeting I record a short voice note. What do I know? What don't I know? What is most important to find out? That is messy input. No strict goal. I think out loud toward the microphone and let AI summarise and structure. Sometimes I land on the right framing. Sometimes the summary reveals that I had the wrong idea about what the meeting was actually about.

After the meeting I record a new note. This is okularbesikning — a term I borrowed from the construction industry, meaning to verify with your own eyes rather than relying on drawings and records. I do not listen to the recording. I do not sit with the transcript. I tell the microphone what actually happened, what I noticed, what deviated from what I expected.

That is still messy input. But retrospective messy. It captures things a transcription never captures — the interpretation of pauses, the feeling that the conversation turned at a particular moment, the thing that was said with just a little too much emphasis.

Then it becomes crisp. Specific follow-ups. Specific corrections. Specific changes. Not a document, but a series of precise edits to the picture that the messy input established.

None of that is automated. And that is the point.

What had no name

I have been writing about voice as raw material since 2023. About how a 15-minute conversation carries more information than a form. About how a client who never writes a word still ends up on a website that sounds like him — precisely because the messy input captured what a requirements document would never have formulated.

But I did not have a word for the calibration. For the active choice to be in messy mode when exploring and crisp mode when specifying. It was a practice without a name.

Knisprigt is closest. It contains the transition — the organic that snaps into form without losing its original voice.

AI translates. That is what it does. It takes messy input and creates knisprigt output. It takes crisp input and applies it precisely. But who decides which mode you are in, and who judges whether the output actually captured the right thing — that is not AI. That is the only job that cannot be delegated.

Why the position is still open

Everyone talks about prompting. About how to write better instructions. About context windows and agents and orchestration.

Nobody talks about calibrating the actual input mode — about training yourself to know when you are in a messy moment and when you need to be knisprigt. That is a human competence. It requires understanding what you are actually trying to accomplish, not just what you write in the prompt.

The position "voice as primary input layer for AI work" has been open for three years. It appears nobody has taken it yet.

That is stranger than it ought to be.

See also: Everything I Say Ends Up as Text (series 2) and From conversation to website (series 9).