Context is everything

In module 1 we agreed that AI is a brilliant new hire who started today and knows nothing about your business. This module is about how you tell it. It is not a module of "magic phrases". It is about what information the model needs to do a task well, how you hand it over, and how you avoid repeating it every day.

We will answer five questions: what context is in practice, why the same question gets different answers, what the AI remembers between conversations, how to build a context document for your business, and when more context stops helping. At the end, the exercise has you do the same task with and without context, which is the fastest way to convince yourself.

What is context when you talk to an AI?

It is everything the model has in view when it writes its answer. Not just your last question: the context includes every message you sent in that conversation, the files you pasted or attached, its own earlier replies, and a set of instructions the app puts in place before you type your first word. Those app instructions are called the system prompt: they tell the model what it is called, what the date is, what tools it has and how to behave. You never see them, but they are there on every turn.

All of that sits inside the context window from module 1, and the practical consequence is simple: if it is not in the window, it does not exist for the model. It does not know your price list because that lives in a spreadsheet; it does not know this customer has already complained twice because that is on WhatsApp; it does not know your tone is casual because you never said so. The reverse also holds: everything that is in the window carries weight, including the paragraph you pasted "just in case" and the wrong answer from ten messages ago.

Why does the same question get different answers?

For two reasons, and it helps to keep them apart. The first you already saw in the module 1 exercise: the model picks each token with an element of chance, the randomness, so two identical runs produce different texts. You cannot control that from the chat, and you do not need to: it changes the wording, and it should not change the substance if the context is good.

The second reason is the one you do control, and the one that weighs most: different context, different answer. "Write me an email to a customer who hasn’t paid" has a thousand correct answers. If the model does not know who you are, who the customer is or what you want to achieve, it picks one of the thousand at random: a generic email, probably in a tone that is not yours, with a threat of legal action you never planned to make. What happens when you add context is that you narrow that range:

• Who you are. "I own a bakery-supplies distributor in Medellín" immediately rules out the bank tone and the lawyer tone.
• The situation. "Eight-year customer, always pays, this time 20 days late and has not answered two messages" completely changes the email that comes out.
• The goal. "I want them to pay this week without feeling scolded" is an instruction; "write me a collection email" is a guessing game.
• The constraints. Four lines at most, no mention of interest, not the word "overdue", signed by me and not by "the team".
• Examples. One earlier email you liked is worth more than three paragraphs describing your tone. In the jargon these are few-shot examples: you show two or three cases and the model copies the pattern.

Notice that none of those five things is a writing technique. They are facts about your business and about this task. That is why module 4 (prompting) comes after this one: the structure of a good prompt matters, but the raw material is context. In 4D terms, this is Description done properly.

Does the AI remember what I told it yesterday?

It depends on the app and its settings, and it matters because many people assume it does and get caught out. The starting point is this: the model remembers nothing on its own. Every new conversation starts with an empty window, except for the system prompt. What you told it yesterday in another chat is not there, unless the app has a memory feature that puts it back into the window.

As of October 2026, according to Anthropic’s help center, Claude has that feature: it saves what it learns about you as a list of topics while you chat, and uses them in new chats. It is on by default on the Free, Pro and Max plans; on Team and Enterprise plans the administrator decides whether it is available and each person turns it on. You can view and edit each topic in Settings, tell Claude in a chat to remember or forget something, and start a conversation with memory off when you do not want it to use what it knows about you. ChatGPT and Gemini have equivalents (Gemini calls it "personal context" and can also look at past chats); the details change often, so check the settings of whichever app you use.

A second kind of "memory" is projects: folders where you upload documents and instructions that apply to every conversation inside that folder. In Claude they are available on every plan (the free plan allows up to five), and each project has its own memory space, separate from the rest. Watch out for a detail that trips people up: inside a project, one chat does not know what happened in another chat of the same project, unless you put it in the project’s documents. Module 6 is dedicated to getting the most out of them.

And there is a third thing that happens inside a single conversation: long conversations degrade. Anthropic says so in its technical documentation: as the number of tokens in the window grows, accuracy and recall go down, and it calls this "context rot". You notice it when, forty messages in, the model repeats a mistake you already corrected, mixes up two customers, or gets more generic with every reply. On top of that, as the conversation nears the limit, the app may start dropping the oldest parts. The fix is not to fight the chat; it is to close it well.

1. 1Ask the model for a summary of the conversation written to be continued in another one (the prompt is below).
2. 2Read it and correct anything wrong or unnecessary. This is a discernment moment: the summary inherits every error you do not catch.
3. 3Open a new chat, paste the summary and carry on. If what you learned is useful for good, do not leave it in the summary: move it to the context document, which is the subject of the next section.

This conversation has gotten long and I am going to continue it in a new chat. Write a summary so that another assistant (or you, with no memory of this) can pick up exactly where we are. Include, in this order:

1. GOAL: what we are trying to achieve, in two lines.
2. KEY CONTEXT: the facts about me, my business and this task that were needed to do the work (no personal data about third parties).
3. DECISIONS MADE: what was settled and why. Include what we ruled out, so it does not get proposed again.
4. CURRENT STATE: what is done, what is half done, what is missing.
5. LATEST VERSION: paste the last approved draft or result in full, not summarized.
6. NEXT STEP: what the next chat should do first.

Do not add information that is not in this conversation. If something was left ambiguous, mark it as pending instead of resolving it yourself.

What is a context document and how do I build one?

If every task needs who you are, what you sell, to whom, in what tone and under which rules, the logical move is to write it once. That is the context document: one or two pages that describe your business the way you would explain it to a new hire on their first day, which you paste (or load into a project) before any task that needs it. It is the most profitable document you will write in this course, and it is the foundation of the prompt library in module 4.

What it usually contains:

• What you do and what you don’t. Products or services in customer language, not catalog language. And what people think you do but you don’t.
• Who you sell to. Your two or three types of customer, what they care about, what worries them, how they find you.
• How you talk. Formal or casual, with or without emojis, words you use and words you avoid. Two or three real snippets of yours that you like, as examples.
• Policies that almost never change. Hours, coverage, delivery times, warranties, cancellations, payment methods, what happens when something goes wrong.
• Prices and facts that almost never change. Only the stable ones. Anything that changes weekly does not go here; it goes in the task of the day.
• What you never say. No promising dates without confirming, no talking about competitors, no discounts in writing, no answering legal questions. This section saves your team more trouble than all the others combined.
• How you decide. The unwritten rules: who gets the exception, when you escalate to the owner, what counts as urgent.

Do not write it from scratch: let the model interview you. Answering questions is far easier than drafting, and a good interviewer pulls out things you would never have thought to include. The prompt below does exactly that, in rounds, and at the end returns the document in a fixed structure so every future version looks the same.

I want to create the context document for my business: the text I will give an AI before any task so it works as if it knew the company. You are going to build it by interviewing me.

Interview rules:
- Ask me questions in rounds of at most 4 questions. Wait for my answers before the next round.
- Start general (what I do, who I sell to) and move to specifics (tone, policies, exceptions, what we never say).
- If one of my answers is vague, ask for a concrete example or a real snippet (a message, an email) instead of accepting it.
- Do 4 to 6 rounds. When you think you have enough, say so and ask me whether anything is missing.

When you finish, write the document with exactly this structure, in plain text, two pages at most:
1. The business in three lines
2. What we offer and what we don’t
3. Our customers
4. How we talk (with 2 or 3 real examples of mine)
5. Stable policies
6. Stable prices and facts
7. What we never say or promise
8. How we decide (unwritten rules and when to escalate)
9. To be confirmed (everything I said with doubts)

To start: I am [YOUR ROLE] at [NAME AND TYPE OF BUSINESS] in [CITY]. Ask the first round.

Where you keep it depends on how you work. If you use an app with projects, you load it into the project’s documents once and every conversation in that project has it. If not, you save it as a text file and paste it at the start of every conversation that matters; it is less elegant and works just as well. If your team uses AI, the document is shared: one version, in a place everyone knows, with one person responsible for keeping it current.

Is more context always better?

No, and this is the idea that separates people who use AI well from people who paste in everything they have. Anthropic, in its guide for people building agents, gives it a name: context engineering, the set of strategies for curating and maintaining the best set of information while the model answers. The principle it proposes fits in one sentence: find the smallest set of high-signal information that maximizes the likelihood of the outcome you want. Context is a finite resource with diminishing returns: every paragraph you add competes for the model’s attention with all the others, and an irrelevant paragraph is not free, it subtracts.

For you, without building anything, that turns into three habits. First, give the information this task needs, not all the information you have: answering a complaint does not require the price history. Second, watch for contradictions: if the document says "we always address customers formally" and the example you pasted is casual, the model will pick one of the two without telling you. Third, clean up: when a conversation has piled up failed attempts and corrections, it is full of noise, and summarizing and starting over beats stacking more on top. The prompt below reviews your context document with those eyes.

This is the context document for my business. Review it as an editor, not a writer: do not rewrite it. Give me three lists:

1. CONTRADICTIONS: statements that conflict with each other or with the included examples (tone, policies, prices, promises). Quote the two conflicting statements.
2. DOESN’T BELONG: information that would not help any typical task (writing customer messages, quoting, answering complaints, summarizing) or that changes so often it should not live here.
3. MISSING: questions a new hire would ask after reading it that the document does not answer.

At the end, tell me in one line which section you would fix first and why.

Document:
[PASTE YOUR CONTEXT DOCUMENT HERE]

The next module, "Claude, ChatGPT, Gemini: how to choose", moves from what to with what: models, plans, what things cost and how to pick a main tool for you and your team. Arrive with your context document ready, because it is the first thing you will load into whichever tool you choose.