Gains:
- Can explain with examples how context determines the accuracy of the output
- Can select and configure what context information is needed for a task
- Recognizes and avoids risks of redundant, contradictory or hidden context
Even if you give the same task to two people, if one of them works knowing the situation and the other does not know anything about it, very different results will occur. The situation is exactly the same for artificial intelligence. Context is information that tells the model "in what background it does this work": for whom, for what purpose, with what history, under what constraints. In this unit we will learn how context determines output, what context to choose for a task, and how too much or the wrong context causes harm.
Why Is Context So Determining?
The model does not know your situation beyond the text you provide. He doesn't know the name of your company, your customer's history, the goal of your project, or what was talked about last week. If you don't provide this information, the model will write with an "average" assumption, and the average will rarely fit your unique situation. Giving context moves the model from “average” to “just for you.”
For example, when you say "write a discount offer to the customer", the model produces a general text. But if you give the context of "this customer has been with us for 3 years, he made a turnover of 400 thousand TL last year, he said that he received a cheaper offer from the rival company and we want to maintain his loyalty", the tone, justification and discount logic of the offer will be completely appropriate.
What Context Does a Task Need?
It is not necessary to write everything for every task; your job is to choose the necessary and sufficient context for that task. These four questions make your job easier:
- For whom? Who the buyer is, his level of knowledge, his relationship with you.
- From where? The purpose of this output is; Persuasion, information or decision?
- With what past? What was talked about before, what happened, what decisions were made?
- With what limits? Budget, duration, brand rules, legal framework.
Setting up context step by step
- Write the task in one sentence.
- Answer the four questions above; Add the answers in bullet points.
- Eliminate unnecessary or contradictory clauses (too much context is also harmful).
- If there is confidential/personal data, mask it (we will return to this topic in unit 11).
- Visually separate the context from the task (title, line, or “Context:” tag).
Little Context / Enough Context / Much Context
Status
Conclusion
little context
It turned out to be general, inaccurate, "suitable for everyone but not for anyone"
Sufficient context
Situation-specific, to-the-point, directly usable output
Excessive/contradictory context
The model confuses the important with the unimportant, obscuring the main demand
Contrary to popular belief, too much context is not good. If you write ten paragraphs of background and say what you want in one vague sentence, the model will drown in the pile of details and miss the main task. Good context is choosing what is relevant; Not spilling everything.
Four Copiable Templates
1) Context block skeleton:
Context:- For whom: [recipient, level, relationship]- Purpose: [what do we want to achieve]- History: [relevant previous events/decisions]- Boundaries: [budget, duration, rule]Task: [single clear verb + output]
2) Customer communication:
Context:- Customer: 2 years corporate, spending ~30 thousand TL per month.- Problem: Last delivery was 5 days late, customer reproached.- Purpose: Maintain the relationship, renew trust. Task: Write a short compensation email (100 words maximum). Constraint: Accept responsibility; don't blame anyone else.
3) Internal report / executive summary:
Context:- Reader: General manager, not interested in technical details, results-oriented.- Purpose: Persuasion for budget approval.- Background: Project delayed by 2 months but scope increased.Task: Write 1 paragraph executive summary + 3-item action list.
4) Continuous context (start of long conversation):
Base all your future answers on this context: - Our company: [domain, size] - Brand voice: [concrete adjectives] - Prohibited: [competitor name, hype, promise] Confirm this and follow these rules in every answer.
Weak Prompt / Strong Prompt
Weak:
Write a response to this complaint we received.[complaint text]
Strong:
Context: Complaining customer 4 years old, loyal; The product is defective, he wants a replacement. Our policy allows unconditional exchange within 14 days and customer on the 6th day. Purpose: quick solution and trust. Task: Write a reply to the customer. Maximum 90 words, give the solution in the first sentence. Restriction: Do not make promises outside of politics; Using accusatory language.Complaint:[complaint text]
The second prompt gives the model the customer profile, policy, duration and purpose; so that the answer is not a "generic apology" but a correct and workable solution.
Three Mini Cases
Case 1 — Real estate consultant. A consultant was producing advertisement texts to "write beautifully". When we added the information "the target buyer is a young family, the proximity of the area to schools should be highlighted, the price is 5% below the market average", the ads directly appealed to the target audience and the search rate per view increased under the consultant's own monitoring.
Case 2 — Project manager. When a PM gave the weekly status email the context of "2 key deliveries missed last week but one risk was closed early; reader senior management," the model struck a balanced tone, neither panicking nor hiding. Without context, the same task would come out either too optimistic or too pessimistic.
Case 3 — HR. When printing a rejection email, an HR professional added the context of “the candidate was strong in the final round, another position may open in the future.” The result was a text that did not close the door or offend the candidate; The version without context was cold and formulaic.
Tip: Use clear labels like “Context:” and “Task:” to separate context from task. In this way, the model does not confuse which part is the background and which part is the request.
Caution: Stop before adding confidential data (real name, TR ID, account number, contract details) to the context. Most of the time, saying "3-year corporate customer" works just as well as giving the real name. We will discuss privacy rules in unit 11.
Common mistakes
- Zero context. Just writing the task and waiting for the model to know the situation.
- Too much context. Piling up ten irrelevant paragraphs and burying the main request.
- Conflicting context. Saying "be formal" and saying "write sincerely and humorously" in the same place.
- Confusing context with task. Explaining the background and not saying clearly what you want.
- Adding confidential data without thinking. Pasting sensitive information without anonymizing.
In summary
- Context is the background information that moves the model from “an average answer” to “an answer specifically for you.”
- Good context answers four questions: for whom, why, with what history, with what boundaries.
- Too little context produces general output, enough context produces accurate output, and too much context produces dispersion.
- Separate the context from the task with clear labels; Choose what is relevant, do not spill everything.
- Anonymize confidential data before adding it; Most of the time the general definition is sufficient.
Application task
Choose a real task from your work (an email, report, or reply). Run the context-free version first. Then answer the four questions (for whom, why, what history, what boundary), add a "Context:" block, and run again. Observe the difference in hits between the two outputs and check for excess context.
checklist
- [ ] I can determine what context is needed for a task with four questions.
- [ ] I separate the context from the task with clear labels.
- [ ] I avoid redundant and contradictory context.
- [ ] I anonymize confidential data before adding it.
- [ ] I can see the difference between context-free and context-free output.