Store-Bought Glaze
Learning Craft With AI Text
Previously: Curation Is Craft the Copyright Office Can See.
In the last post I wrote about curation as a creative act the law can actually see: the selection, the arrangement, the human fingerprints on the assembled work. That argument was about legal recognition. This one is about what happens at your desk, inside the sentences, before anyone else sees it. It’s about craft.
When I was studying ceramics, learning to glaze was its own semester-long problem. You could buy glaze premixed in a bottle (cone six satin, iron red, commercial celadon) and it worked fine. Predictable. You brushed it on and you got what the label promised.
The problem was that you got exactly what the label promised.
Every other student with the same bottle got the same result. The glaze didn’t respond to your clay body, your silica content, your firing temperature. It had no relationship to the specific piece it was on. Instructors called it “store-bought” in a tone that said: not wrong, but not yet yours.
Making your own glaze meant something different. You started from raw materials: feldspar, silica, whiting, maybe iron oxide for color, bentonite to hold the mixture in suspension. You measured. You tested on tiles. You fired once and adjusted, fired again. The same base recipe behaved differently in oxidation than in reduction, differently on a stoneware body than on porcelain. Understanding why required understanding the chemistry, or at least the craft-knowledge equivalent of it: what each material contributed and what would happen when you changed the ratio.
That process had a name. We called it learning the material.
Working with LLM text is the same problem. The system has its own defaults, and they’re persistent. The defaults exist because humans produced them, not because the model is broken. Bullet lists everywhere. The “not X but Y” contrast move, used so often it has become a tic. Three-part rhythm: first point, second point, third point, brief synthesis. Em-dashes as load-bearing glue between ideas that haven’t quite earned adjacency. Whether these patterns are baked into the model by training or are just internet writing habits the models inherited, I don’t know — but they’re persistent and they’re recognizable.
Every one of those patterns appears in human writing. The model learned them from us. That’s actually the point: the model is doing what it was trained to do, optimizing for patterns that read, statistically, as coherent and helpful. When you get a response built entirely from those moves, you’ve received something that technically works. Predictable. Every other person asking a similar question gets a structurally similar answer.
This is store-bought glaze.
There’s an older precedent for the accept-or-reject loop, one that doesn’t feel threatening at all because we’ve been using it long enough that nobody argues about it anymore.
Grammar checkers. Spell check. Grammarly. Microsoft Word’s grammar checker, going back decades.
For years, every time you typed a sentence in Word, a colored underline could appear. Red for spelling, green (then blue) for grammar. The software would suggest a revision. You looked at the suggestion. Sometimes you took it. Often you didn’t, because the corrected version was technically grammatical but wrong for your rhythm, your tone, your specific effect. The sentence was supposed to start with “And.” The comma splice was intentional. The fragment landed harder that way.
Nobody claimed grammar check made you less of a writer. Nobody argued the resulting text belonged to Microsoft. The writer’s judgment, the sustained series of small accept-or-reject decisions, was obviously the creative act. The tool surfaced options. The person chose.
Predictive text on phones extended this further once smart keyboards became standard. You’re mid-sentence and the keyboard offers three words. Most of the time you tap past them. Occasionally one is exactly right and you’re mildly surprised. Occasionally one reveals that you were about to write something imprecise and you revise.
Working with an LLM is that same loop, expanded by an order of magnitude. The completions are longer. The suggestions are more developed. The number of accept-or-reject decisions per paragraph goes up. But the structure of the act, the human in the loop, judging, is not new.
Where it becomes craft is in what you refuse.
Letting an LLM write your paragraphs without intervention gets you the default glaze. The prose functions. It’s organized. It’s not yours. The sentence lengths march in approximate lockstep. The transitions use the same four connectors. Every concept gets its explanatory clause. The rhythm is median, calibrated to scan smoothly for a hypothetical reader who wants to absorb information without friction.
If your goal is information transfer with minimum friction, that result may serve you well. But if your goal is a prose voice that sounds like you, with your habits of observation, your particular way of pausing before a point, your instinct for which comparisons illuminate and which just pad, then you have to break the defaults.
This is editorial work, not mystical intervention. You read what came back and you notice: this sentence moves too fast, I’d slow down here. This list should be two sentences. This contrast is the “not X but Y” move and I just did it two paragraphs ago. This em-dash is holding together two thoughts that should be two sentences with space between them.
Ceramicists talk about fitting the glaze to the clay body. The glaze and the clay expand and contract at different rates during firing; if the fit is wrong you get crawling, crazing, or shivering — the glaze pulls apart, or cracks, or flakes. A glaze that works on one clay body fails on another. It has to be calibrated to the specific piece it will live on.
Voice is the same calibration problem. The corrections you make to LLM output are the adjustments. They’re not cleanup; they’re the craft.
None of this happens automatically, and it doesn’t happen fast. The first time you override a default, you might not be sure why you’re doing it, only that something feels off. That instinct is worth following. It develops. After enough repetitions you can name what’s wrong: the rhythm was three-part when the idea needed two; the bullet list fragmented something that gains force from continuous pressure; the “importantly” before the sentence was throat-clearing for a thought that would’ve been stronger without the warning label.
Experienced use of grammar check looked the same way. At first you accepted suggestions without much scrutiny. Eventually you developed a feel for when the software’s notion of correctness diverged from your actual intent. You accepted fewer suggestions. You understood them well enough to explain why you were rejecting them. That judgment was craft.
Augmented authoring with an LLM follows the same arc, compressed or extended depending on how seriously you treat the practice. What you’re building is editorial instinct calibrated to a new tool. The tool has tendencies. You learn them. You decide which ones serve you.
There’s still something left unaddressed, and I want to name it before the next post takes it up. Craft this deliberate, this clearly yours, can still produce work with no aura.
Aura is the other problem. It’s harder.
A ceramic artist can buy bagged clay, use store-bought glaze, fire in an electric kiln, produce technically accomplished pots, and sell them. That’s craft. Whether those pots are art depends on something the technique alone doesn’t supply. The work that follows this one is about why AI content, even well-crafted AI content, so often falls flat, and what the missing element is.