The 120-Page Contract vs the Window

The supplier contract runs to 120 dense pages — about 90,000 words — and Tessa needs one thing out of it: what the cancellation terms are for group bookings. Not a summary, not an overview. One answer, which is either right or expensive.

She could paste the whole thing, if her fingers and the product allowed it. She should not, and this page is why. A real document meeting a finite window is the collision this book has been walking toward since Chapter 1, and the craft of winning it is a skill Tessa will use every week for the rest of her career.

Think of how you would ask a colleague the same question. You would not read them the contract cover to cover and then ask. You would find the two sections that plausibly cover cancellations, read those aloud, and ask about those. That instinct is correct, it is entirely portable, and everything below is that instinct made deliberate.

Do the Token Maths First

Chapter 1 left a rule of thumb for exactly this moment: a token is about three-quarters of a word. Turn it around and 90,000 words is roughly 120,000 tokens. That is the weight of this document in the model's own units, and it took ten seconds of napkin arithmetic to find out.

Whether 120,000 tokens fits depends on the product, and windows have been growing. But fits is the wrong target, because barely fitting is almost as bad as not fitting. Attention over an enormous context is uneven: material at the very start and the very end is used more reliably than material buried in the middle, so a clause on page 63 of a giant paste is the likeliest thing in the document to be quietly underweighted. A document that fits can still be poorly read.

So the maths is not there to tell you whether the paste will be accepted. It is there to tell you what kind of job you are facing. Under a few thousand tokens, paste it and get on with your afternoon. In the tens of thousands, start choosing. At 120,000, choosing is the whole task.

Cut to What Is Relevant

The first and best move is to stop treating the contract as one object. Contracts have structure — a table of contents, numbered sections, headings written by people who wanted things to be findable. Tessa opens the PDF, searches for "cancellation", finds section 9.4 and a cross-reference in the definitions, and pastes those three pages into the chat with her question.

Three pages of the right text beat 120 pages of everything, and they beat it twice over. The answer is better, because everything in view is relevant and nothing has to be sifted for. And the ask is cheaper — a fact worth nothing today, while she is using the browser chat box, and worth real money in Chapter 7, where every token in is billed.

This is the same move as Chapter 2's context habit, scaled up. There, the skill was pasting the paragraph the model could not know. Here, the skill is finding the paragraph in the first place — which is a research skill, not an AI skill, and one Tessa already had.

When You Cannot Guess Where It Lives

Sometimes the search does not work. The contract calls it "termination of group reservations", the word "cancellation" appears forty times in irrelevant places, and Tessa has no idea which sections matter. Then go piecewise.

Paste a chunk — a chapter, twenty pages, whatever fits comfortably — and ask a small, cheap question of it: does this section say anything about cancelling group bookings? Yes or no, and quote it if yes. Then the next chunk, in a fresh chat so the previous one is not still crowding the window. Collect the hits in your own notes, then paste the hits together and ask the real question against them.

Two rounds of small questions and one round of the real question — that is divide and conquer, and it is more work than pasting the file. It is also the difference between an answer that came from the relevant clause and an answer that came from wherever the model's attention happened to land.

Summaries Lose Things, and You Choose Which

There is a shortcut that looks clever and fails in a specific way: summarize the contract first, then ask the summary. It fails because a summary is not a miniature of the document. It is generated text about the document — plausible, usually fair on the gist, and produced by a machine that was choosing what to keep without knowing what Tessa was going to ask.

For "what kind of agreement is this, broadly?" a summary is fine. For "how many days' notice do we need for a group of twelve?" it is a trap, because the exact clause is precisely the sort of detail a summary drops, and its absence looks like nothing at all. Nothing in the answer will say a clause went missing.

So for any answer someone will act on, insist that the model works from the actual text — and then ask it to quote the clause it relied on. That request costs one line and changes the kind of answer you get. An unverifiable claim becomes a claim with an address: Tessa reads the quoted words, finds them in her own copy, and decides for herself whether they say what the model said they say. Chapter 3 taught this as a habit for facts. Here it is the habit for documents, and it is the single most useful sentence in the chapter.