A prompt is the input you send a foundation model to get a response, and most prompts are built from a small set of parts: an instruction, context, input data, a description of the output you want and, for image models, a negative prompt. This lesson takes a prompt apart, so you can tell which part to fix when a response goes wrong, a skill the exam tests through short scenarios.
The instruction
The instruction is the task itself: "Summarize this review in one sentence", "Classify this email as personal or commercial", "Translate this paragraph into Spanish." It's the one part almost every prompt needs. AWS's prompt guidance stresses keeping it simple, clear and complete. If a classification task should pick from fixed labels, the instruction should name those labels, or the model will invent its own wording. AWS also suggests placing the question or instruction near the end of the prompt, after any long text, so the model knows what to do with everything it has just read.
Context and input data
Context is background that helps the model interpret the task, such as "The following text is from a customer support chat" or "You are answering questions for a bank's employees." It narrows the model's focus and steers its tone and vocabulary.
Input data is the material the model works on: the review to summarize, the email to classify, the document to answer questions about. Context and input are easy to confuse. A quick test: input data is what the instruction acts on, while context tells the model how to read it. Many prompts also include demonstration examples, which the next lesson on prompting techniques covers.
Output indicators
An output indicator describes the shape of the answer: its length, format or structure. "In one phrase", "as a bulleted list", "return only the year" and "wrap each name in name tags" are all output indicators. Without one, a model often returns a free-form answer that's longer or looser than you need, which is a problem when another program has to read the result. AWS's examples show that adding an output indicator alone can turn a rambling summary into the short phrase the user wanted.
Negative prompts
A negative prompt lists what you don't want in the result. It matters most for image generation. Stability AI's Stable Image Ultra in Amazon Bedrock, for example, has a separate negative prompt field: you describe the scene you want in the main prompt and list unwanted elements, such as "clouds, text, people", in the negative prompt. Image models handle words like "no" or "without" in the main prompt unreliably and may even draw the thing you mentioned. That's why exclusions go in the negative prompt instead. Text models have no separate field, but you can still state what to avoid ("don't mention competitors") as part of the instruction.
| Part | What it does | Example |
|---|---|---|
| Instruction | States the task | "Summarize the review in one sentence" |
| Context | Frames how to read the task | "This text is from a hotel review site" |
| Input data | The material to work on | The review itself |
| Output indicator | Sets the answer's form | "Return only a star rating from 1 to 5" |
| Negative prompt | Excludes unwanted content | "blurry, watermark, text" |
On the exam
Expect a symptom and a choice of fixes. If the answer's content is right but its format is wrong, add an output indicator. If the model misreads what the text is, add context. If the model does the wrong task, sharpen the instruction. If an image keeps including something unwanted, use a negative prompt rather than writing "no ..." in the main prompt. And don't jump to fine-tuning when a clearer prompt would fix the problem.