Give It Better Context (The Model Only Knows What Is on the Desk)

A client writes: “Our supplier delivered six weeks late. Can we terminate the contract and recover our costs?” You paste the sentence into an AI assistant. The answer comes back in ten seconds: fluent, confident and useless. It applies the law of the wrong country, assumes a notice was sent, and never mentions the amendment that moved the delivery date.

The model did not fail. It answered the question it was given. The model only knows what is on the desk at the moment it answers. Your sentence was the only thing on the desk.

This lesson takes that one client email through five briefs, from the empty desk to a brief that makes the assistant stop and ask the right question. Then it gives the same brief to an agent. The method is the same in a general assistant, a legal platform or a tool your firm built. Only the buttons differ.

Everything on the desk is context

Most people learn the word prompt. The useful word is context: everything the model can see when it writes its reply. The prompt is one item on the desk. The others are usually larger.

• Your message. The question and the facts you typed.
• Hidden instructions from the product. The vendor wrote them. Lesson 4.1 calls this the harness.
• The earlier turns of the conversation. Sent again with each message, as lesson 4.2 showed.
• The documents you attach. Read in full when short, searched when long.
• Passages that a search tool brought back. From your files, a legal database or the web. Lesson 2.2 explains what comes back and what does not.
• The examples you give. A past letter, a model clause, a filled template.

Engineers at Anthropic, which makes Claude, define context as “the set of tokens included when sampling” from the model, and describe the work as finding “the smallest possible set of high-signal tokens that maximize the likelihood of some desired outcome” Anthropic · Effective context engineering, Sep 2025 (opens in a new tab). They call it context engineering rather than prompt engineering, because the wording of the question matters less than what sits around it.

The illustration above shows the proportions. In a normal exchange about a contract, the message you typed is the smallest item on the desk, and most of the rest arrived without you choosing it. Better results come from choosing it.

Five briefs, one question

Here is the client's email, asked five ways. Each step puts one more kind of context on the desk. Watch what the answer stops guessing.

Three things happen across the five steps.

• The guesses go down. The first answer invents a legal system, a notice and a loss. The second has one guess left and says so. From the third, every sentence points to a clause.
• The order of the advice changes. The typed brief gets the law right and still gives the wrong next step, because the contract caps damages for delay unless the agreement is terminated. Only the document could say that.
• At the last step, the model stops. Given an example and permission to ask, it asks the two questions a senior lawyer would ask before drafting: why was the supplier late, and does the client want the goods or the money.

Nothing changed in the model between step one and step five. The desk changed.

The model fills every gap silently

The first answer is the dangerous one, and not because it is wrong. It is dangerous because nothing in it says what the model did not know. It never writes “you did not tell me the jurisdiction”. It picks one, usually the most common in its training data, and writes as if you had said it.

Researchers at Carnegie Mellon measured this in 2026. When a prompt left a requirement unstated, models met it by default about 41% of the time. Prompts with unstated requirements were also twice as likely to break when the model or the prompt changed, sometimes with accuracy drops of more than 20 points ACL · What Prompts Don't Say, 2026 (opens in a new tab). A brief that works by luck today stops working when the product updates its model, as lesson 4.1 explains.

So read the tone of an answer as a signal. A confident answer with no caveats is a sign of an empty desk, not of certainty. The more an assistant hedges and asks, the more it has noticed what it is missing.

The model also mirrors your framing. “Confirm that we can terminate” produces a memo that confirms. “Assess whether we can terminate, and give the strongest argument against” produces an assessment. Lesson 2.3 explains why. Here, the point is practical: the frame of your question is context too, and the model takes it as given.

What a brief contains

The brief at step five is fifteen sentences. It has five parts. None of them needs legal English.

The situation is what a colleague would already know: who you act for, the governing law, the dates. Add today's date. The model does not know it, and limitation periods and deadlines go wrong without it.

What the client wants is the question behind the question. “Can we terminate?” has a different answer depending on whether the client wants out, wants the goods and compensation, or wants leverage for a renegotiation. Lawyers extract this from clients in the first five minutes and then forget to pass it on.

What is settled stops the hedging. If you do not say which facts are agreed, the model treats all of them as open and wraps every sentence in a condition. Three short sentences remove three paragraphs of “if” and “assuming”.

The limits include one rule most briefs forget: which source wins when two conflict. If your summary says one delivery date and the amendment says another, the model may follow either. Say that the signed documents govern. Then say who will read the result, and how long it may be.

What done looks like shapes the thinking, not only the layout. Ask for a conclusion and you get a conclusion with reasons attached. Ask for the issues, then the facts each one needs, then the view, and the model works through the problem in that order.

The usual advice is to brief a model as you would brief a junior. That analogy is useful, and it breaks in one place that matters. A junior remembers last week, knows the client and can walk to your office. The model has none of that. Brief it as you would a brilliant temp who started this morning, in another country. Everything a colleague would already know has to be on the desk.

Examples say more than adjectives

“Professional”, “concise” and “client-friendly” carry almost no information. One letter you sent last year carries a hundred preferences you would never think to write down: how long, how formal, how you address the other side, how much law to cite, where the demand sits, how the deadline is phrased.

The engineers who build these systems say the same. Rather than long lists of rules, the Anthropic guide recommends “a set of diverse, canonical examples that effectively portray the expected behavior”, and adds: “For an LLM, examples are the ‘pictures’ worth a thousand words” Anthropic · Effective context engineering, Sep 2025 (opens in a new tab).

In practice:

• One or two good examples, not ten. More examples cost tokens and blur the pattern.
• Say what the example is for. “Match its structure and tone” is different from “use its arguments”. Say which.
• Say what to change. Names, dates, amounts and the governing clause will differ. Say so, or the model may copy them.
• Choose the example carefully. A model matches what it sees. A weak letter as the example produces a polished weak letter. Anonymise what must be anonymised, as lesson 3.4 explains, but keep the relationships between parties, dates and amounts consistent, because the answer depends on them.

Show one good piece of your work rather than describing what good looks like.

More is not better

The natural reaction to “give it more context” is to attach everything. The research says the opposite. A 2023 study at Stanford found that models answered best when the relevant passage sat at the beginning or the end of a long input, and “significantly” worse when it sat in the middle TACL · Lost in the Middle, 2024 (opens in a new tab). In 2025, Chroma tested 18 models and found that performance fell as the input grew, even on simple tasks, and that “even a single distractor reduces performance”: one passage that looks like the answer but is not Chroma · Context Rot, Jul 2025 (opens in a new tab). Anthropic's engineers describe the same thing as an “attention budget” that every token spends down.

So the deal folder is worse than the contract and the amendment. Fifty pages of correspondence about other matters is worse than nothing. The skill is selecting and pointing, not uploading. Name the clauses that matter, as the third brief did. For a bundle too large to read, lesson 2.2 explains how the search works and how to steer it.

Three smaller habits follow from the same idea.

• Check what the tool received. Text, Word files and clean PDFs arrive intact. A scanned PDF without text recognition arrives as nothing. Tracked changes and comments are often dropped. Wide tables come through scrambled. If the document matters, ask the assistant to quote the first lines of the clause you care about, and see what it saw.
• Start a new conversation when the task changes. Every earlier turn is still on the desk, with its assumptions and corrections. A long chat about one matter shapes the answer to the next.
• Remember that context is sent again at each turn. On a subscription this spends your allowance; on the API it is on the invoice (lesson 4.2). Pointing to a clause is cheaper and more accurate than pasting the contract six times.

Ask for an answer you can check

Lesson 1.5 showed that verification is where the saved time goes. The brief decides how much of it there is. Three requests turn checking from rereading into spot-checking:

1. Quote the clauses you rely on, with their numbers. You check five quotations instead of reading thirty-four pages.
2. List what you had to assume. The list is a map of your own brief's gaps. It often shows you a fact you forgot you knew.
3. Say what you could not find. An answer built on clause 15 gives no hint that a side letter exists. A sentence that says “I found no side letter” does.

And one request beats all three: ask before you draft. The fifth brief ends with it, and the model used it to ask about the cause of the delay, the one fact that decided the case. The best brief ends with permission to ask.

The same brief, given to an agent

In a chat, you put things on the desk. An agent fills its own desk: it opens the folder, chooses which file to read, decides when it has enough and drafts (lesson 1.2). That changes what the brief has to do. It must now say where to look, what counts as the right version, and when to stop.

With the thin brief, the agent opened the first file that matched, read the original delivery date, and drafted a confident letter on a delay twice the real one. Nobody told it that amendments exist. Nothing in the finished letter shows the error.

With the full brief, the same agent in the same folder found the amendment and the side letter, computed the right delay, and stopped on the fact that decides the case. The difference is four sentences in the brief: look for amendments and side letters, the signed documents govern, check the notice rule, and stop before anything goes out.

For an agent, the brief decides what goes on the desk. Context quality moves from “what did I attach” to “what did I tell it to find, to trust, and to check”. The five parts are the same. The cost of leaving one out is higher, because nobody is reading between the steps.

Where this lives in your tool

Every product has its own names for these mechanisms. The mechanisms themselves are the same seven, whatever the logo.

Two of these deserve a word. Saved instructions are the situation and the limits written once, so that every conversation starts with them on the desk. Connected sources are what let the fourth brief find a side letter you had forgotten. Lesson 1.3 explains both, and lesson 6.2 shows how to turn a brief you keep rewriting into one.

Eight mistakes

• No jurisdiction, no dates. The model picks both and tells you about neither.
• No reader. A memo for a partner, a note for a client and an email to the other side are three different documents. Say which.
• A document with no question. contract.pdf and “thoughts?” produces a summary of everything and an answer to nothing.
• The grand preamble. “You are the world's best commercial lawyer” adds no fact to the desk. One sentence about the client does more.
• Everything in one chat. The morning's problem shapes the afternoon's answer. New task, new conversation.
• The output asked for last, or not at all. Say what done looks like before the model starts, not after it has finished.
• The whole folder. 34 pages that matter beat 400 pages that might.
• Trusting a summary of a long document without pointing to the clause. If the search missed it, the summary is complete and wrong, and reads exactly like one that is complete and right.

What to remember

• The model only knows what is on the desk. Context is everything it can see, and your message is usually the smallest part.
• It fills every gap silently. Confidence without caveats signals missing context, not certainty.
• A brief has five parts. The situation, what the client wants, what is settled, the limits, and what done looks like. Fifteen sentences are enough.
• Examples beat adjectives. One good past document carries more than a paragraph of instructions.
• Selection beats volume. The right documents, pointed to, beat all the documents. Long inputs and look-alike passages make answers worse.
• End with permission to ask. Quotes with clause numbers, a list of assumptions, and questions before drafting turn checking into spot-checking.
• For an agent, the brief decides what goes on the desk. Say where to look, which source governs, and when to stop.

You have now written this brief five times. Next, lesson 6.2 shows how to write it once, as saved instructions and skills your whole team can use.