Context: The Subtractive Discipline

The Intent article covered the first lever of prompt discipline: naming what the model must produce, in a sentence a three-year-old could act on. This is the deep dive on the second lever, and it is the one that decides what you pay.

Context is where token bills quietly compound. It is also where accuracy is most often lost — not because the model was weak, but because it was drowned. Every extra line you paste into a prompt is a line the model must read, weigh, and bill for, on every call, forever, until you take it out.

That is why the discipline is subtractive. The goal is not “give the model everything we know.” It is “give the model exactly what is required to produce the asked-for output.” Everything else is noise.

Why context is where the cost lives

Most teams treat context as a setup cost. You paste the policy in once, the model has it, done. That is not how it works. Context is a per-call cost. Every prompt sends the same context again, and the token bill compounds by usage, not by build effort. What looked cheap when a handful of people were using it becomes the line item on the quarterly review when it is being used at scale.

There is a second, quieter cost: accuracy. Long context windows do not automatically produce good answers. They produce answers that must first be pulled out of surrounding noise. The larger the noise, the harder the extraction, and the more often the model settles for something that sounds right rather than something that is.

A disciplined prompt does not solve this by paying for a bigger context window. It solves it by sending less.

The half test

Before any prompt with context goes into production, I run one test.

Could I remove half of this context and still get the right answer?

Not “could a smaller context work in general” — a specific yes-or-no on the actual passage in the actual prompt. Delete the second half of the pasted policy. Send it. Does the answer still hold?

If yes, the second half was noise. Delete it in production. Then run the test again on what remains.

If no, ask the follow-up question: which specific lines did the answer depend on? Keep those. Delete the rest. That is the excerpt the prompt actually needed. Everything else was you paying, on every call, for context the model was not using.

This sounds destructive. It is meant to. Subtraction is the reflex the discipline is trying to build. When a prompt is failing on cost or accuracy, the first move is not to add — it is to cut, then check. The prompts that survive the half test are the ones that get on with it.

What context actually contains

A disciplined context answers three questions.

The source. What is the authoritative document, data, or excerpt the answer must draw from? If the answer can come from more than one source, name which one wins. If the sources contradict, decide which the model should trust. Ambiguity in source is ambiguity in answer.

The scope. How much of that source is actually relevant? A two-paragraph excerpt beats a forty-page document every time. If you cannot narrow the source by hand, use retrieval — but the goal of retrieval is still the smallest excerpt that answers the question, not the biggest excerpt you can afford.

The fallback. What should the model do when the context does not contain the answer? Silence is a decision. So is “I don’t know.” So is “not covered in the retrieved policy.” A prompt without a fallback instruction produces confident wrong answers when the context runs out — which is a worse failure mode than a “don’t know” you can act on.

Miss the source and the model averages across contradictions. Miss the scope and you pay for irrelevance on every call. Miss the fallback and you get answers that sound right and are not. The three cases below are composites — realistic in shape, drawn from patterns I have seen more than once, not tied to any one client — and each one is a single miss.

Case study 1 — The kitchen sink

An HR policy assistant, built in Copilot Studio, answering questions about leave, benefits, and expense rules across a large workforce.

BEFORE:  {Attach the full 40-page HR handbook.}
         Answer the employee's question below using the handbook.

It worked, in a sense. The answers were technically correct. But the token cost per resolved query ran close to three times the estimate, and the CFO’s first quarterly bill turned into a project.

Source was named. Fallback was implicit — good enough. The miss was scope. The handbook contained the answer, but thirty-eight pages of it were not relevant to any single question. The model read them all, on every call, to find the two paragraphs that were.

AFTER:   {RAG retrieval selects the two policy clauses most
          relevant to the question.}
         Using only the retrieved clauses below, answer the
         employee's question quoted at the end.

Token cost dropped to a third. Answer quality went up, because the model was no longer being distracted by thirty-eight pages of irrelevance. The one-time cost was setting up retrieval. The ongoing cost fell every day thereafter.

Case study 2 — The wrong document

An internal customer-support Copilot, drawing on a knowledge library to answer product questions.

BEFORE:  {Retrieve all documents matching the customer's
          question from the knowledge library.}
         Answer using the retrieved content.

The retrieval was working. The answers were also inconsistent — the same question, asked twice, produced different answers depending on which documents the retriever surfaced. Support agents started routing questions around it.

Scope was there. Fallback was there. The miss was source. The knowledge library contained three generations of the same policy — old, new, and superseded — with no signal to the model about which was current. The retriever pulled all three. The model averaged them, and the averages were wrong.

AFTER:   {Retrieve from the canonical, version-controlled
          knowledge index — deprecated documents filtered
          out.}
         Answer using only the retrieved content. Cite the
         source document ID in your response.

A single canonical source, an index that excluded deprecated versions, and a citation instruction so any wrong answer could be traced back to the document that produced it. Consistency returned the same week the index was cleaned.

Case study 3 — The confident hallucination

An internal Copilot in a regulated context, answering questions about specific rules from a curated knowledge base.

BEFORE:  Using the rules below, answer the user's question.

Source was named. Scope was tight. The miss was fallback. When the user asked a question the rules did not cover, the model still answered — plausibly, confidently, and sometimes wrong. In a regulated context, that is not a productivity feature. That is a compliance incident in slow motion.

AFTER:   Using only the rules below, answer the user's
          question. If the answer is not explicitly present
          in the rules, reply exactly: "Not covered in the
          retrieved rules — please escalate to a human
          reviewer." Do not infer.

The fallback closed the gap. The model still answered inside the rules. Outside them, it said so, in a form the downstream flow could branch on. Confident hallucinations stopped.

Write the source and the scope first

The practical discipline is an ordering one. Before you paste any context, decide three things — in this order.

First, name the source. One document, one index. If more than one, name which wins. Write that in the prompt, not in a comment somewhere else.

Second, narrow the scope. If you can name the specific passage by hand, paste that. If you cannot, use retrieval — but treat retrieval as a question of “the smallest excerpt that still answers the question,” not “the largest excerpt I can fit.”

Third, decide the fallback. What should the model say when the context runs out? Write that too.

If any of the three refuses to come — you cannot name the source, or you cannot narrow the scope, or you cannot write the fallback — that is the prompt telling you the underlying question has not been decided. Do not paper over it with more context. Go and decide it.

Quick check

Before this prompt ships, ask one question: if I removed half of this context, would the answer still be correct?

If yes, delete the half that was not doing work, and run the check again on what remains. If no, name the specific lines the answer depended on, keep those, and let the rest go. The prompt that survives that reflex is a prompt that costs less and answers better, on every call, for as long as it runs.

Context is the second of five levers. The next, Boundaries, is where the discipline moves from what the prompt sends to what the prompt returns — and where the same subtractive reflex applies to the shape of the answer.

Where to start

If you have not read the framework this lever belongs to, start there: it covers all five levers and includes a ten-point scorecard you can run any prompt through.

→ Read: Prompt Discipline: The Five Levers Behind Production-Ready Enterprise AI

→ Read: Intent: The First Sentence That Decides What Your Prompt Costs — the deep dive on Lever 1.

→ Read: The Agent That Could Not Leave Pilot — the governance teardown that sits alongside these discipline pieces.

→ Download: The Prompt Discipline Card Pack (PDF) — five printable one-page cards, one per lever. Card 2 is the Context companion for this article; Cards 3–5 are stubs waiting for their deep-dive articles.