Text a Program Can Read

Tessa's bookings tracker has an import button, and the note beside it says it accepts structured text. She asks the colleague who set the tracker up what that means, and he says: JSON. She braces for programming.

What she finds is labelled boxes. JSON is a strict way of writing down information so that each piece is labelled with what it is, leaving a program nothing to guess at. There is no logic in it, nothing that runs, nothing to execute — it is data wearing name tags. The letters are an abbreviation from the programming language it grew out of, and knowing that helps nobody, so treat JSON as a word in its own right. Reading it takes about ten minutes, and this page is those ten minutes.

What Goes Wrong with Ordinary Sentences

Start with the problem, because the solution looks arbitrary until you have felt it. Tessa writes: the Lakeland tour, four guests, June 12th. Every human at Waymark reads that correctly and instantly.

Now be a program reading the same line. Which words are the tour name — "Lakeland", or "the Lakeland tour"? Is "four" a count of guests or a rating out of five? Is "June 12th" this year or next? The information is all there, and none of it is labelled, so the program has to work out which part is which from the way the sentence happens to be written. Write the next one slightly differently and the guessing starts over.

Think of a suitcase with a handwritten note inside saying where it should go. A person who opens it can route the bag. A luggage tag works differently: fixed fields, name and address and flight, on the outside, so any handler at any airport routes it without opening anything or reading anything they have to interpret. Labelled fields are what make handling automatic.

Reading It, Gently

Here is Tessa's booking with labels on. Read it as three pairs: a label, then the value that belongs to it.

{"tour": "Lakeland", "guests": 4, "date": "2026-06-12"}

Curly braces wrap one item — one booking, here — and everything inside them describes that one thing. Inside, each label sits in quotes, then a colon, then its value, and a comma separates one pair from the next. The label tour holds Lakeland, the label guests holds four, the label date holds a date written the same way every time. Nothing is left to interpretation, which is the entire point.

Values come in two flavours you will notice straight away. Text sits in quotes; numbers do not. That is not decoration — it is how the program knows that four guests is a quantity it can add up, while a tour name is a piece of text it should not try arithmetic on. There is also a plain true-or-false value, written without quotes, for fields that are answers to yes-or-no questions.

Second example: more than one booking. A list is written in square brackets, with the items inside it separated by commas.

[
  {"tour": "Lakeland", "guests": 4, "date": "2026-06-12"},
  {"tour": "North Shore", "guests": 2, "date": "2026-06-19"}
]

Square brackets for a list of things, curly braces for one thing, labels naming every value inside. That is the whole reading skill, and it does not get harder — a big file is this same pattern repeated two thousand times, sometimes with a set of braces tucked inside another set.

Third example, and this one is Tessa's actual work: a classified review, in the shape she will ask for on the next page.

{"review_id": 1184, "tour": "Lakeland", "rating": 5,
 "category": "praise", "refund_mentioned": false}

Five labels, five values: which review this is, which tour it concerns, the rating as a number, which of Waymark's four buckets it went into, and whether it mentioned a refund. The buckets are the ones she defined in Chapter 4; the labels are hers too. Nobody handed her this shape — she decided what the tracker needed and wrote the labels down.

Asking the Model for It

Getting this out of the chat box is one sentence plus one example, which by now should sound familiar. "Return each review as JSON with the fields review id, tour, rating, category, and refund mentioned" — then paste one filled example, exactly as Chapter 2 taught, and the shape comes back reliably.

Models produce this format fluently, because they have read enormous amounts of it. Fluent is not the same as guaranteed, though, and the guarantee is what strictness costs you. A stray sentence of commentary before the data, a missing comma, a quote mark that never closes — and the import fails. Not partly: the tracker rejects the file and says nothing useful about why.

That cuts both ways, and the second edge is in your favour. Because the format is strict, a program can tell instantly whether a file is valid — no judgement, no reading. A machine that refuses malformed input is a machine that tells you when something went wrong, which is more than prose ever does. The page after next makes a habit of exactly that.

One practical note, so a rejected file does not become a mystery. Ask for the data and nothing else — no introduction, no closing remark, no explanation of what it did. The model's instinct is to be conversational, and a friendly opening line is the single most common reason a perfectly good file will not import.

Why This Shape Turns Up Everywhere

The reason to spend ten minutes on this is not the tracker. It is that this is how programs hand labelled information to each other across the whole internet, all day, in numbers nobody can picture. Your phone's weather, a payment going through, a flight search — labelled fields moving between machines, in this format or a close relative of it.

Which is why the tracker imports it, why half the output worth automating is worth automating because it comes in this shape, and why what Tessa learned in ten minutes is about to be worth far more than the afternoon it saved. When Chapter 7 lifts the lid on the chat box and shows the request underneath, what is inside that request is JSON — labels and values, curly braces, a list. She will recognize it on sight.

And there is a nearer payoff. A shape a program can read is a shape a program can check: ratings that must be between one and five, a category that must be one of four words, a field that must not be empty. Judgement is not required for any of that, which means it need not be Tessa's judgement doing it. Hold that thought; the last page of this chapter picks it up, and Chapter 8 finishes it.