Prompting as context design
LLMs are powerful, but they are not situated. They do not automatically know your goal, your audience, your data, your constraints, or your definition of success. Working well with AI means supplying that missing context.
The human supplies the situation
An LLM may know a great deal in general, but it does not know the specific situation unless you provide it. In most useful AI work, the human contributes three things:
| What the human provides | Example |
|---|---|
| Goal | “Help me decide which vendor to recommend.” |
| Context | “The buyer cares most about reliability, implementation time, and total cost.” |
| Success criteria | “Produce a one-page recommendation memo for a non-technical manager.” |
This is why working with an LLM is not just asking a question. It is giving the model enough of the situation to produce a useful answer.
Prompting is context design
Common prompt-engineering advice is really advice about context. “Give it a role,” “be specific,” “include examples,” and “describe the output format” all help the model understand the situation and produce something useful.
| Prompting advice | What it provides | Example |
|---|---|---|
| Give it a role | Perspective | “Act as a marketing analyst.” |
| Be specific | Task boundaries | “Compare these vendors on cost, speed, and risk.” |
| Give examples | Pattern | “Use the format shown in this sample memo.” |
| Describe the output | Deliverable | “Return a table with recommendation, evidence, and caveats.” |
The context tradeoff
Context is powerful, but providing the right amount of context is key. Too little context causes the model to guess. Too much or poorly organized context can cause the model to miss what matters. The goal is to provide the right context at the right time.
| Problem | What happens | Better approach |
|---|---|---|
| Too little context | The model fills gaps with its own assumptions. | State the goal, audience, constraints, and desired output. |
| Too much context | The model may miss important details or drift toward the wrong task. | Provide the most relevant information and label it clearly. |
| Poorly placed context | The model may underweight key instructions buried in the middle. | Put the task and key constraints where they are easy to find. |
Example: too little context versus useful context
Too little context:
Write a recommendation.Useful context:
Act as a business analyst. I am comparing three vendors for a
student-run food delivery project. The decision criteria are cost,
reliability, implementation speed, and ease of use. Write a one-page
recommendation for a non-technical manager. Start with the recommended
vendor, then give three reasons and two risks.The second prompt is better because it gives the model the goal, role, decision criteria, audience, format, and success criteria.
Context Format: Providing Examples
Due to LLM's unique ability to recognize patterns, providing examples can be more effective than describing your requirements.
Suppose you want the model to summarize customer feedback in a specific format. A less useful prompt might be:
Summarize this customer feedback clearly and include the main point,
evidence, and business implication.A more useful prompt shows the model the pattern:
Summarize each customer comment using this format:
Theme: Delivery reliability
Evidence: "My order arrived 40 minutes late, and no one updated me."
Business implication: Customers may tolerate occasional delays, but
lack of communication makes the experience feel worse.
Now use the same format for the comments below.The second prompt works better because it gives the model an example of the desired structure, level of detail, and style. Instead of only describing the requirement, the human shows the model what a successful answer looks like.
A Note on Context Window
An LLM can only use information that fits inside its context window: the amount of text the model can consider at one time, including the user’s prompt, prior conversation, uploaded material, and the model’s own response. The size of this window depends on the model. For example, many Claude models support roughly 200,000 tokens, while some newer long-context models from Anthropic and OpenAI support up to about one million tokens.
A larger context window does not mean every detail is used equally well. Research on long-context language models finds that models can be less reliable when the relevant information is buried in the middle of a long prompt. This pattern is often called being lost in the middle.
Exercise
Choose a simple business task, such as drafting a customer email, comparing two products, or recommending a vendor. Write two prompts for the same task:
| Prompt version | What to include |
|---|---|
| Version 1 | A short prompt with very little context. |
| Version 2 | The same request with a goal, audience, constraints, and desired output format. |
Run both prompts in an LLM. Compare the outputs. What did the model assume when context was missing? What improved when context was provided?