Last reviewed:
What is a prompt? Definition and business implications
A prompt is the natural-language instruction sent to an AI model to obtain a response. Its quality directly conditions the quality, relevance, and cost of the generated output. A good prompt is not long; it is precise, structured, and contextualised.
The prompt is the cornerstone of interaction with an LLM. Concretely, it is the text you type (or that your application sends in the background) to query the model. A minimal prompt can fit in one sentence; an elaborate prompt can include a role assigned to the model, a business context, examples (few-shot), a precise instruction, an expected output format, and negative constraints (“do not include X”). The quality of a prompt is measured against three criteria. Precision: no ambiguity about the expected output. Contextualisation: the model has the business elements it needs to answer. Structure: clean sections, clear hierarchy, instructions without contradiction. A bad prompt produces generic, off-topic responses, or confident hallucinations. A good prompt produces specific, actionable, reproducible responses. The same question, badly framed then well framed, can yield two radically different answers, with no change of model.
Concrete example
Illustrative case. A salesperson at an industrial SME writes responses to calls for tender. Prompt 1 (minimalist): “Write me a response to this tender: [text]”. Output: generic, flat, no personalisation. Prompt 2 (structured): “You are sales lead at a French SME manufacturing machine tools, 80 staff, based in the Auvergne-Rhône-Alpes region. Reply to the tender below following this plan: (1) understanding the need, (2) three technical differentiators, (3) after-sales support, (4) delivery schedule. Cite our regional assets. Professional tone, no superlatives”. The second version generally yields a more specific output that is easier to use.
FAQ
What is a prompt?
It is the natural-language instruction sent to an AI model to obtain a response: the text you type, or that your application sends in the background. Its quality directly conditions the quality, relevance and cost of the generated output.
Is a good prompt necessarily long?
No. A good prompt is not long, it is precise, structured and contextualised. Its quality is measured against three criteria: precision (no ambiguity about the expected output), contextualisation (the model has the business elements it needs) and structure (clean sections, instructions without contradiction).
Why does the same question yield such different answers?
Because everything depends on the wording. The same question, badly then well framed, can yield two radically different answers, with no change of model. A bad prompt produces generic, off-topic responses or confident hallucinations; a good prompt produces specific, actionable ones.
Should prompts be managed as a company asset?
Yes. Prompts used in production are a documentary asset, on a par with your procedures. They deserve to be versioned, shared and tested, because prompting skill is acquired, not bought.
What impact does a prompt have on cost?
Every prompt sent is billed in tokens: at a constant rate, a prompt twice as long costs twice as much in input tokens. And it does not necessarily improve the output, because past a certain density added context interferes more than it helps. Sobriety is therefore a direct economic lever.
See also
Further reading
Sources
- Prompt engineering overview, Anthropic documentation 2026. https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overview
- The Prompt Report: A Systematic Survey of Prompt Engineering Techniques, Schulhoff et al., arXiv:2406.06608, 2024 (v6 2025). https://arxiv.org/abs/2406.06608