Writing prompts: four steps to follow

‍

Writing prompts, i.e. instructions given to generative Artificial Intelligences, cannot be improvised. To be effective, you need to follow a well-defined process. We take stock.

What if knowing how to ask questions was a skill, and in some cases even a technique? The question was once raised when the first search engines appeared on the Internet. Today, it's just as acute when we come into contact with the generativeartificial intelligence embedded in intelligence tools. Knowing how to ask the right questions to Chat GPT, Midjourney or Google Bard is not something you can improvise, but a process in which each stage has its own importance. Arnaud Marquant, Director of Operations at KB Crawl SAS, explains.

Four steps to effective prompts

Step 1: formulate the main question rigorously

Writing prompts, i.e. giving instructions to generative AIs, requires first and foremost the definition of an objective. What exactly are we asking of artificial intelligence? The key here is to be as clear as possible, so as toassign the AI a role that can be applied as effectively as possible. To achieve this, you need tobe fairly prescriptive about the function assigned to Chat GPT, for example. Details such as "You're a watchdog based in France and you work in the insurance sector" will enable the AI to express itself with a certain level of language. Conversely, something like "you work for a bank" will be too vague...

Step 2: Context

The second step is to put the command into context, again, as clearly as possible. This action is unique in that it enables the AI to better structure its response. In this case, it's a question of detailing what is expected of the generative artificial intelligence: asummary of articles, an abstract... It's advisable tospecify the language level, the tone of the text, the target, the precise objectives. It is also possible to ask the AI to produce its editorial proposal in an accessible or, on the contrary, technical format. When writing a text document, for example, it's also important to mention who the readers will be (the marketing team, the management team, etc.).

_____________________________________________________________

Keep up to date with the latest marketing news by subscribing to our newsletter.

_________________________________________________________________

Step 3: Format and tone

Artificial Intelligences such as Chat GPT strictly apply what is asked of them. For example, it's important to specify the nature and scope of the desired deliverable: a summary, a paragraph structure, a reasoning, an analysis... It's also possible to ask it to express itself in the form of a bulleted list, a table or... a visual for certain special cases. Example of an order: "You're going to write a summary, starting with three catchphrases, continuing with a list of the topics covered in these articles in the form of dashes, and writing a conclusion". Note that the tone can be more specific. It can be formal, educational, technical...

Step 4: Define boundaries and invite AI to make revisions

It's important for generative AIs to be aware of the limits of their action. A formulation such as "write a rather long review article" is far too vague. It is preferable to use wording such as "write an article of less than 2,500 characters, including spaces", or "don't write sentences of more than 20 words". It's also time to get back to the drawing board. The first requests made to generative AI must imperatively be the subject of new questions, which will make it possible to adjust and refine the initial request. Continuing the conversation with the artificial intelligence in this way enablesnew elements to be added, or others to be subtracted. It's an iterative process.

Today, we're in the early stages of a relationship: in a few months' time, we human beings will have acquired the methodological reflexes needed touse generative artificial intelligence with maximum efficiency. In particular, we'll have learned to be wary of the use of adjectives, which lead AI down the emotional slope and produce unsatisfactory results. We'll also know how to adopt a neutral, directive and precise approach, like a high school teacher giving his pupil an oral test for the baccalauréat. Will the teacher and student be able to stay in their respective places, without the Chat GPT student taking the teacher's place? Most certainly. The prompt exercise shows us how, in the end, the machine simply bends to the demands of human intelligence.

‍