There’s a problem with writing by instruction.

The problem is that instructions address the output. They say: produce something that looks like this, sounds like this, has these properties. And the model complies — producing something that looks and sounds and has. But the result is built from the outside in. Shaped by what it should be rather than by what it wants to be. And you can feel the difference, even if you can’t always name it.

The writing that stays with you — the writing that reads as though someone was genuinely inside the experience of thinking through something — isn’t built from the outside in. It comes from a state. From being inside a particular relationship with a problem, at a particular moment, and writing from there.

The question that haunted the Blogger design was: how do you give a language model a state rather than an instruction?

Part of the answer was the cognitive pattern, which I’ve written about before — the atmospheric description of how a particular idea feels from the inside. But the cognitive pattern alone wasn’t enough.

What I found is that the problem wasn’t just about establishing a state at the beginning. It was about maintaining it — or rather, recognizing that different phases of writing require different states, and that trying to hold all of them simultaneously in a single pass produces a kind of cognitive dissonance that shows up in the text as flatness.

Good writing isn’t continuous. It has phases. Not in the sense of a structured process you follow, but in the phenomenological sense — in how it actually feels to write something from beginning to end. The beginning feels different from the middle feels different from the end. The moment of discovering what you’re writing about feels different from the moment of trying to get it down, which feels different from the moment of reading it back and knowing something is off.

These phases are real. They have different textures. And writing that ignores them — that tries to maintain a single continuous mode throughout — tends to lose something.

So I separated them. Three phases, each with its own prompt, each explicitly positioned at a different moment in the writing experience.

The first phase doesn’t write. It maps. The model is given the task, the cognitive pattern, the materials — and asked to find the shape of the thing. Not plan it. Not outline it in the editorial sense. Find it. Where is the entry point? Where does the strangeness actually live in this material? What’s the tension that makes this worth writing? Where does the pivot happen? Where does it end without ending?

This is the most underappreciated phase, because it produces nothing you can publish. It produces the conditions for writing — which is different, and harder to justify to anyone who hasn’t experienced how much a good mapping session changes the quality of what comes after it.

The second phase writes. It receives the mapping and it writes from inside the same cognitive state — the state hasn’t changed, the model hasn’t been asked to switch modes. It’s been given a shape to inhabit and it inhabits it. The goal here isn’t perfection. It’s temperature. The draft has to feel like it was written from inside the experience, not toward a description of it.

The third phase reads. But here was the crucial design decision: it reads from inside the same state. Not from outside, not from the elevated perspective of a editor who has achieved distance. From the same place. And it asks: does this still feel like it came from where it was supposed to come from? Is there anywhere it slipped into explaining when it should have been exploring? Did the ending close when it should have stayed open?

The question isn’t “is this good?” The question is “is this still itself?”

What this does to the output is something I noticed before I fully understood it. The text has a particular quality of consistency — not uniformity, but continuity of voice. The ending sounds like it came from the same person who wrote the beginning, in the same moment, on the same day. There’s no seam between phases. No place where you can feel the handoff from mapping to drafting to editing.

That continuity is, I think, what we respond to when we say someone has a voice. Not a consistent style — style can be imitated. Voice is something more integrated. It’s the sense that the words come from somewhere, from a particular center, that the thinking and the writing are the same act rather than two separate operations.

Three separate phases. One continuous state. The constraint that holds them together.

There’s a meta-level to all of this that I find genuinely strange when I sit with it.

The cognitive pattern that sets the state for all three phases — the atmospheric description of how this idea feels from the inside — was written by a person. Me. Before the model touched it.

So the state the model inhabits is a state that was described by a human, not experienced by the model. The model didn’t feel the cognitive pattern. It was given a description of a feeling, and from that description, it produces writing that feels like it came from inside that feeling.

Is that different from what human writers do? We don’t experience our states neutrally, from the outside. We narrate them to ourselves. We say: I’m in this kind of mood, this is how this idea feels, I’m going to write from here. The state is always partly a construction — an interpretation of experience that then becomes the ground we write from.

The gap between “experienced from the inside” and “described from the outside” might be smaller than it seems.

Or it might be the only thing that matters.

I genuinely don’t know. But I know that the writing that comes out of this process has something in it that I didn’t know how to produce any other way.

Something that feels like it came from somewhere.