A prompt is the instruction, question, context, or constraint supplied to a generative AI model to steer its output. It can be a quick request, but useful prompts usually make the task, audience, format, and relevant background plain. The model may survive typos and vague wording, yet guessing your intent is a terrible workflow to depend on. section 2.13 section 6.1 section 6.0
Treat prompting as an iterative conversation rather than a magic spell. Try a request, inspect the result, then refine the instruction when it misses the mark. Useful controls include assigning a role, requesting a response format, asking for step-by-step help, adding context, and using system messages or reference text to guide behavior. section 6.1 section 6.7
A practical prompt is usually:
- Specific about scope, detail, length, audience, or desired result. section 6.3
- Clear about the order and format of the response, especially when extracting or organizing information. section 6.4
- Positive about the action wanted, rather than relying on a negative instruction the model might mishandle. section 6.7
- Grounded in supplied context when the answer must reflect particular documents, facts, or constraints. section 2.13 section 3.4
Turning a broad request into a usable one
“Tell me about quantum mechanics” leaves the job wide open. “Explain quantum mechanics in 200 words or less and in terms a 12-year-old can understand” supplies a length and audience. Likewise, an email request improves when it names the intended tone and the evidence supporting a raise. section 6.3
Prompting against your own documents
A prompt can query an index built from local documents: ask whether events conflict across a set of Sherlock Holmes stories, then inspect the answer independently. A second request for a new story can specify characters, ambience, style, and plot, but generated prose may echo source text too closely. Add an explicit no-verbatim-quotations constraint and verify both accuracy and originality. section 3.4
Prompts are not limited to chat. They can drive API calls and automation, where model settings such as temperature, presence penalties, and Best Of also affect results. For image generation, describe the subject and add an adjective and artistic style; extra color or texture details can further constrain the image. section 2.14 section 4.2
|
|
Tell me about quantum mechanics. |
|
|
Explain quantum mechanics in 200 words or less and in terms that can be understood by a 12-year-old. |
Compare a generic request with a constrained one to see how audience and word-count requirements turn a topic into an actionable prompt. section 6.3
A polished-looking completion is still only a completion. Check important claims independently, particularly when a prompt asks the model to interpret documents, draft legal language, or generate code. Domain knowledge helps you ask better questions and catch confident nonsense before it escapes into the wild. section 3.4 section 3.5
