Context Is a Purchase

A model knows Vue, TypeScript, SQL, and the shape of a thousand REST APIs. It does not know your database, your route table, or what your team named anything. Every task that touches a system the model has never seen splits along that line: general knowledge it already has, and local facts it cannot possibly have. Context is how you sell it the local facts — and every token of it is bought, whether you priced it or not.

What the model cannot know

Ask for a query against a schema the model has never seen and it will not stop to say so. It will write the query anyway, with column names that sound right — amount, created_at, user_id — because plausible invention is what a model does with a gap. The result is fluent, well-formatted, and does not run.

That is the failure mode that makes under-buying invisible: missing context does not produce an error message, it produces confidence. Nothing in the output marks the line where knowledge ran out and invention began. You cannot see the seam by reading. You can only know where it has to be: everything local is on the far side of it.

When the wrong guess runs

Sometimes invention is worse than wrong — it is runnable. Ask a model to cancel order ord_4821 against an API it has never read and it will guess the RESTful default: DELETE /api/orders/ord_4821. On a real API that endpoint can exist, return 200, and do the wrong thing — delete a draft, say, when live orders are cancelled through POST /api/orders/:id/cancel. Nothing about the response says "wrong endpoint".

The fact that prevents this costs almost nothing. A route table — one line per route, a couple hundred tokens — states the endpoint outright. The full OpenAPI dump states it too, buried in four thousand tokens of schemas.

await fetch(`/api/orders/${id}/`, {
  method: '',
})

Name the fact, then buy the source that states it

Before reaching for anything, ask what a competent person would need in front of them to do the task. Not what would be nice to have open in a tab — what is required. For "total order value per customer over the last thirty days, highest first", the required fact is which tables and columns exist. That is the whole list.

Now price the menu. Say these four things are available:

The table definitions — orders(id, customer_id, total_cents, placed_at, status) — answer the question at 400 tokens. The manual is fifteen times the price and answers a different question: how the tool works, which the model already knows from training. The decisive source is the cheapest one here, and that is not a coincidence: it is the most specific one. You are not buying knowledge — it has knowledge. You are buying the local delta.

select customer_id, (total_cents) as total_value
from orders
where placed_at >= now() - interval '30 days'
group by 
order by total_value desc

More context is not a safer bet

The instinct under uncertainty is to hand over everything — schema, rows, manual, logs — and let the model sort it out. That habit has two costs, and both are quiet.

The first is the bill. Everything on the menu above costs sixteen times the necessary price, and context is priced per call: a "just in case" source that rides along in an agent loop is a tax on every iteration. Over-buying has no visible failure, so the two have to be measured separately.

The second cost is quality. Irrelevant context is not neutral padding; it is material the model will try to use. Hand over the Pinia docs alongside a request for raw SQL and the answer drifts toward the manual's idioms. Every unnecessary source is another surface the model can anchor on, and it does not tell you when it has.

Where this bites

Pasting the whole SFC to ask about one computed. The rest of the file is anchor material, billed on every message that follows. Excerpt the function and the types it touches.

Reading a confident answer as proof the context was sufficient. Fluency is what invention looks like too. Check the one fact that had to come from context against the source you bought, or run the result.

Buying documentation to answer a local question. Manuals describe the tool; the tool is general knowledge. If the question is about your system, the answer lives in an artifact of your system.

Letting a "just in case" source ride along. Paying sixteen times once is trivia. Paying it on every call of a loop that runs all day is the line item that makes someone audit the prompt.