What is a prompt?

What it means at work

A prompt can be one line or several pages. It can include the task, background, the text to work on, examples, and the format you want back. The more the tool knows about the job, the less it has to guess. In practice, most disappointing AI output traces back to a prompt that left out who the audience is, what the source material is, or what a good result looks like.

What the sources say

• Definition. Schulhoff et al. say developers and end-users interact with generative AI systems through prompting and prompt engineering.
• A prompt can hold examples. Brown et al. write that humans can generally perform a new language task from only a few examples or from simple instructions.
• Where you put things matters. Liu et al. found that performance is often highest when relevant information occurs at the beginning or end of the input context.
• Be specific. The Anthropic guide says to be specific about the output format and constraints you want.
• Test it on a colleague. The Anthropic guide suggests showing your prompt to a colleague with minimal context on the task and asking them to follow it. If they would be confused, the model will be too.
• Prompts sit in a limited space. The Anthropic glossary defines the context window as the amount of text a language model can look back on and reference. Very long prompts and attachments compete for that room; see context window.

An example

Weak prompt: ‘Write a follow-up email.’ Stronger prompt: ‘Write a follow-up email to a client who missed Thursday’s call. Two short paragraphs, friendly but direct, offering two new times next week. Here is the last email we exchanged: …’ The second gives the tool a reader, a goal, a length and real context.

A common misconception

Many people think prompting needs magic phrases. It mostly needs the same things you would tell a capable new colleague: the goal, the background and what done looks like. If the first answer misses, say what to change instead of starting over.

Sources

• Schulhoff et al., “The Prompt Report: A Systematic Survey of Prompt Engineering Techniques” (arXiv): “Developers and end-users interact with these systems through the use of prompting and prompt engineering”
• Brown et al., “Language Models are Few-Shot Learners” (arXiv): “humans can generally perform a new language task from only a few examples or from simple instructions”
• Liu et al., “Lost in the Middle: How Language Models Use Long Contexts” (arXiv): “performance is often highest when relevant information occurs at the beginning or end of the input context”
• Anthropic, Prompting best practices: “Be specific about the output format and constraints you want”
• Anthropic, Prompting best practices: “Show your prompt to a colleague with minimal context on the task and ask them to follow it”
• Anthropic, Claude Platform Docs glossary: “refers to the amount of text a language model can look back on and reference”