Context Design & Why It Matters

Introduction

Module 2 ended on the limit of its own technique: a prompt can carry a good role, a precise instruction and a well-chosen example, and still produce the wrong answer — because the assistant was never told something it had no way of knowing.

That is what this module is about. It is the second half of the skill, and it is the half most people never learn: better prompting is not only about asking better, it is also about supplying the right information.

You will leave able to ask a different question of a disappointing answer than the one Module 2 taught. Not "which component should I change?" but "what did this need to know that I never gave it?" — and, just as importantly, "what should I not have shared at all?"

By the end of this module you will be able to look at a task and say what the assistant needs to know, what would only get in the way, and what should never leave your hands.

Context Is What the Task Needs to Know

Module 2 drew this line briefly and deferred the depth to here. The distinction fits in two sentences. Instruction — what you want done. Context — what the assistant needs to know in order to do it well.

"Summarize these replies for our volunteer committee" is an instruction. That the committee is five people, that two joined last month, and that they are deciding whether to move the event date — that is context. The instruction stays the same across every version of that request; the context is what changes the answer.

Context tends to be one of a small number of things:

• •Who the result is for, and what they already know.
• •The situation the task sits inside.
• •Background that explains why the work is being done.
• •Constraints that any acceptable answer has to respect.
• •Source material the answer should be drawn from.
• •Preferences and conventions you have already settled.
• •Current facts, supplied by you, that the assistant has no way to look up.

What It Knows, and What It Does Not

An assistant does not automatically have access to your world. Unless you supplied it in the conversation, it does not have:

• •Your documents, files, notes, or messages.
• •Your organization's processes, decisions, or internal terms.
• •Your teacher's rubric, your employer's policy, your group's rules.
• •Your customer's history or your colleague's opinion.
• •What was said in a meeting you attended.
• •Local or current conditions where you are.
• •Which version of a document is in front of you.
• •Preferences you have never written down.

If the answer depends on information the assistant was never given, better wording alone cannot reach it. That sentence is this module's spine. A learner who takes it seriously stops rewriting the same prompt five times and starts asking what is missing. Some larger systems can go and fetch relevant information automatically rather than waiting for a person to paste it — that is a real idea, and it is Module 6's. Here, the supplier of context is you.

Relevant Context, Not More Context

More context is not better context. Useful context is relevant, specific, current enough for the task, organized, and proportionate. Everything else carries a cost, and the costs are concrete:

• •It buries the facts that matter among facts that do not.
• •It introduces details that conflict with each other, leaving the assistant to choose.
• •It anchors the answer to whatever was most prominent rather than most important.
• •It exposes information that never needed to leave your hands.
• •It makes the result harder for you to check, because now there is more to verify.

A useful test: for each thing you are about to include, can you say in one sentence how it would change a good answer? If not, leave it out and see whether the result suffers.

Organizing Context So It Can Be Used

When a request carries more than a couple of facts, separating them helps — for the assistant, and for you, because you can see what you actually supplied. A simple shape:

• •Task — what you want done.
• •Background — the situation, briefly.
• •Important facts — the specific things that must be right.
• •Constraints — limits any acceptable answer has to respect.
• •Source material — the text the answer should be drawn from.
• •Desired result — the form the output should take.

These labels are organizational aids, not special commands. No assistant is watching for the word "Background", and none of this is a syntax you must reproduce — plain paragraphs work too. The point is that instructions, reference material and constraints stay distinguishable, because when they run together the assistant has to guess which sentence was a rule and which was a fact. Keep source material clearly marked as material: a paragraph you pasted for the assistant to work from is not an instruction to obey, and mixing the two is the most common way a request quietly changes meaning.

Give the Smallest Source Material That Still Works

You can paste an excerpt or a summary of something a task depends on. The beginner rule: supply the smallest amount of source material that still contains what the task needs.

Two cautions worth stating plainly. First, do not assume the assistant can open a file, follow a link, or see a document unless the tool you are using genuinely put it there — if you did not supply it, it is not available. Second, a summary you wrote is itself a claim; where the detail matters, supply the detail rather than your paraphrase of it.

Finding What Was Missing

When an answer disappoints, Module 2 taught you to ask which component to change. This module adds the other question: what did this need to know that I never supplied?

The usual gaps are few enough to check quickly — wrong audience assumed, wrong situation assumed, wrong time period, missing source material, missing constraints, or missing local knowledge about how your group or organization actually does things.

• •Right shape, wrong assumptions → you left out the situation.
• •Right situation, wrong emphasis → you left out who it is for.
• •Confidently states something untrue about your work → it never had the source material.
• •Ignores a rule you take for granted → you never wrote the constraint down.

Context Is Not the Same as Truth

Supplying context does not make an answer correct. It changes what the assistant is working from, which is worth a great deal — but the working material can itself be wrong:

• •Wrong context — a fact you believed and did not check.
• •Outdated context — accurate last quarter, misleading now.
• •Biased or partial context — one side of a situation presented as the whole.
• •Incomplete context — the half of the story you happened to have.

Any of these produces a fluent, confident, well-grounded-sounding answer that is wrong. Grounding an answer in supplied material does not remove invented detail, and it does not remove your responsibility to check. You verify the source, the relevance, the currency where it matters, and the result.

Context Has a Privacy Cost

The context that helps most is often the context you should share least. Before pasting, ask four questions: Does this task actually need this? Can a summary do instead? Can I remove the names and identifying details? Am I allowed to share this at all?

Student records, customer information, employee details, health information, financial records, confidential business material and other people's private lives all tend to arrive disguised as helpful background. Most of the time a task needs the shape of the situation, not the identities inside it — and a version with names replaced by "a member" or "the customer" works just as well.

Deeper prompt safety, privacy and prompt-injection awareness belong to Module 7.

The Two Shapes Worth Remembering

Instruction + relevant context → a better-grounded result → your check → your decision.

Better-grounded means the assistant had more of what the task required. It does not mean correct, and the check does not become optional because the context improved.

Everything you could share → filter → only what this task needs → share safely.

The filter is the skill. Most people supply either far too little, because they assume the assistant already knows, or far too much, because pasting everything feels safer than choosing. Both produce worse answers than a considered middle.

Writing for a Real Audience, in Any Language

Context often includes who the audience is — and that is where care is needed. Useful audience context is specific: "for parents of children in their first year at the school, most of whom have not used this system before" tells the assistant something it can act on. "Write it for people in country X" is too broad to help, and invites a stereotype rather than a description.

Language, locale, units, date conventions and tone expectations are all legitimate context. The techniques here are language-neutral: supply the context in whichever language you think most precisely in, where the assistant supports it, and the reasoning is unchanged. Deeper multilingual technique is Module 8.

Human + AI

You choose the context. You decide what is safe to share. You judge whether the answer fits reality. An assistant can work only with what it has been given — it cannot know what you left out, and it cannot tell you that something it never received would have changed its answer.

• •The assistant does not know your organization, and does not remember everything.
• •Context does not guarantee accuracy, and grounding does not eliminate invented detail.
• •More context is not always better; there is no perfect context.