Read the story first. The thing it is teaching you will not say its name until the end. That is the whole point.
You heard right, and it was Amanda Askell. She is the philosopher who leads personality and values work on Claude at Anthropic, and in a conversation on prompting she shared the technique she reaches for most when she wants to actually understand something new.
Her favorite prompt, almost word for word:
Three moves, in a deliberate order: story first, reveal second, plain explanation third. She does not let you know the topic up front, because if you know the label you stop paying attention to the thing and start filing it. Hold the label back, and your mind has to do the real work: build the shape of the idea from the inside, then snap it to a name only once the shape is already yours.
Why it works is not mysticism. We remember stories far better than we remember exposition. Almost everything an AI was trained on is exposition. So when you ask for a parable, you are asking the model to hand you a format that sticks better than most of what it ever read, on demand, for any topic you name. Askell says she now carries these stories in her head for fields she never formally studied.
So here is one, built for you, on the topic you asked about. Do not skip to the bottom.
Old Beatrix had been blind since birth, and she was the finest nose in the province. Merchants came from three rivers away with stoppered vials, asking the same question: what is this, and what is it worth?
She never answered by naming. She answered by walking.
Her workshop was an enormous round hall, and over sixty years she had filled it with jars, ten thousand of them, on shelves that ran floor to ceiling and wall to wall. To a visitor it looked like chaos. There was no alphabet to it, no order by price or by region or by the year of the pressing. But Beatrix moved through it without a candle, because the room was not arranged by name. It was arranged by likeness.
When a new scent came to her she would unstop it, breathe, and carry it slowly into the hall, and her feet would find the place it belonged. The deep resinous ones lived together near the north wall. The bright citrus ones gathered by the door where the draft was. And in a quiet hollow in the middle, where the light never reached, sat all the scents that made a person think of grief: funeral lilies, cold churches, rain on old stone. None of them smelled alike to a chemist. All of them belonged to the same sunken corner of her room.
A boy apprenticed to her once asked how she could find a jar she had placed forty years before. "I do not remember where I put it," she said. "I remember what it is near. Bring me something that smells of an orchard after frost, and I will walk you to within an arm's reach of every frost-orchard scent I have ever owned, even the ones I met only once."
This was her real gift, and it took the boy years to see it. She had turned smelling into standing somewhere. Two perfumes no one would ever describe with the same words, one called "widow's veil" and one called "the empty pew," stood close enough to touch, because she had placed them not by their words but by what they did to a person. And two perfumes with nearly the same name, "morning rose" and "mourning rose," sat at opposite ends of the hall, because one opened the chest and one closed it.
When she grew old the merchants feared her knowledge would die with her. So the boy, a man now, did a patient thing. He walked the whole hall with a measuring cord, and for every single jar he wrote down its exact place: so many paces north, so many east, so high on the shelf. He turned her entire room into a long ledger of positions, a list of numbers, one row for each scent.
And then something happened that Beatrix herself had never been able to do. With the ledger the man could calculate. He could ask which two scents sat nearest each other in the whole hall. He could ask which jars formed a tight little crowd leaning together, and what they shared that no one had ever named. He could take a brand new vial, set it down once, write its numbers, and find its ten closest cousins without walking a step. He could even ask the strangest question of all: what would sit exactly between the smell of a newborn's scalp and the smell of fresh-cut hay? The numbers pointed to a spot on a shelf, and on that shelf was the warm, milky, grassy thing that has no name in any language but that every mother knows.
The man understood at last what his teacher had done her whole life. She had been quietly insisting that meaning has a place: that anything which means something can be set down in a room so that things meaning the same thing end up as neighbors, and the distance between any two things is the difference between them, made walkable.
And here is the thing he built, and the reason I am telling you this.
He never owned the perfumes. He owned a recording of a hundred thousand conversations, the calls your team records and then mostly forgets.
Each one he carried into a hall with far more directions than north and east and up. A hall with hundreds of directions, more than a body could ever walk, that only a machine can move through. And he set each conversation down by what it meant, not by the words it used. So every call where a buyer quietly lost faith ended up in the same sunken hollow, whether they said "I need to think about it," or "let me loop in my team," or said nothing revealing at all. Every call that felt like the deals you eventually won leaned together in one bright crowd near the door. And a new call, set down once, told you in an instant which crowd it had joined.
He had turned listening into standing somewhere. He had given your conversations a place.
The concept was vector analysis of transcripts. Beatrix's hall is a vector space, and her nose was an embedding.
| In the hall | In the machine |
|---|---|
| A jar of scent | A piece of transcript: a sentence, a moment, or a whole call |
| Beatrix's nose, judging by what a scent does to a person, not its words | The embedding model: it reads meaning, not surface words |
| The exact place a jar sits in the hall | The vector, a list of numbers that are its coordinates |
| North, east, up, and directions with no name | Dimensions: often hundreds or thousands of them |
| Things that mean the same end up as neighbors | Semantic similarity, measured as distance (cosine distance) |
| "widow's veil" and "the empty pew" touching | Different words, same meaning, so the vectors land near each other |
| "morning rose" and "mourning rose" at opposite ends | Similar words, opposite meaning, so the vectors land far apart |
| The sunken grief corner | A cluster: a theme that forms on its own, unlabeled |
| "Walk you to within an arm's reach" | Semantic search / nearest-neighbor lookup |
| The measuring-cord ledger of positions | Vectorizing: turning text into numbers you can compute on |
| "What sits between a scalp and fresh hay?" | Vector arithmetic: meaning becomes addable and subtractable |
| Placing a brand new vial instantly | Embedding a new call at the moment it lands |
| The bright crowd of winning-deal calls | Classification by likeness to past winners |
A computer cannot compare meanings. It can only compare numbers. An embedding model is a trained reader that takes any chunk of text and returns a long list of numbers, a vector, which you can picture as a single point in a space with many directions. The trick that makes the whole field work is one property: the model places text so that similar meaning lands in nearby positions, even when the words are completely different. Meaning becomes geometry. Once it is geometry, you can do arithmetic on it.
From that one move, everything else falls out:
similarity how close are two points, so how alike in meaning are two moments. search given one point, find the nearest others, so find the calls most like this one. clustering let crowds form on their own, so discover the recurring themes nobody pre-labeled. classification is this new point nearer the "won" crowd or the "stalled" crowd. arithmetic add and subtract directions, so move from one meaning toward another on purpose.
A bridge to your other life: this is the same move a clinician makes recognizing a presentation. You do not match every symptom against a list. You feel that this case sits near ones you have seen before, and that nearness is the diagnosis forming. Embeddings are that instinct, written down as coordinates so a machine can have it too.
SPICED is a hall.
Every field, Situation, Pain, Impact, Critical Event, Decision, is a region of meaning. A call is not "a Pain moment" because someone said the word pain. It is a Pain moment because, in the space, it sits near every other Pain moment you have ever recorded. Tagging a transcript against SPICED is the machine asking Beatrix's question: which corner of the hall does this belong in?
That is what turns SPICED from a methodology a human has to remember into a protocol two machines can speak. When a buyer's AI and a seller's AI need to agree on what just happened in a conversation, they do not trade your slide definitions. They trade positions in a shared hall. The semantic layer you keep describing is, underneath, exactly this: conversations given a place, so that meaning can be compared without words getting in the way.
So the thing you wanted to learn and the thing you are trying to build are the same thing seen from two ends. You did not learn a tool today. You looked at the floor of your own cathedral.
A tiny, illustrative hall. Each dot is a real-sounding call moment, placed by meaning. Drop a new line in and watch which crowd it joins. The coordinates here are hand-placed to show the idea, not computed by a live model.
The lesson under the lesson: you now have a reusable way to learn almost anything, the one Askell uses. Paste this into me, or into any Claude, and name your domain at the very end.
Now go carry a story in your head.