To excel on ChatGPT and write your best queries, OpenAI provides its prompt guide

OpenAI has released a guide to help users write better prompts when using generative artificial intelligence. Goal: Get what you want from the models quickly, without having to repeat the query over and over.

ChatGPT, Gemini, Copilot… Generative AIs have taken the web by storm and are used by a growing number of people. On the other hand, it is sometimes difficult to make the model understand exactly what you want, and that is normal.

Prompt engineering, a discipline which consists of refining and optimizing interactions between humans and generative AI systems, is today a particularly sought-after profession. With a few tips, it is nevertheless possible to take better advantage of AI without being a real expert.

Write the clearest instructions possible

First of all, OpenAI recommends writing clear instructions so that the model is better able to provide an answer in line with the initial query. For example, by specifying the length of the desired answer, asking for a certain level of expertise, or even showing an example of what is desired.

It is also advisable to divide complex tasks into simpler subtasks. “ Complex tasks can often be redefined as a flow of simpler tasks in which the results of previous tasks are used to construct the inputs of subsequent tasks », Details OpenAI. A prompt specifying the tasks to be accomplished one by one rather than a general request should make the model much more efficient.

Also, the company explains that it is necessary to give the AI ​​time to “ to reflect on “. Indeed, it is possible to ask a model to reason on its own for certain queries, such as knowing whether the solution to a mathematics problem is correct. The model can also be led to study a question in more depth if it is asked to check if it has not forgotten a step in its reasoning, for example.

Do not hesitate to check the results with other tools

For even more technical queries, OpenAI recommends using third-party tools. It is thus possible to compensate for certain shortcomings of a model by indicating to it the results obtained through another platform. The use of code to perform more precise calculations as well as calling on external APIs can also be considered.

In order to avoid errors, it is also advisable to test changes systematically. “ In some cases, changing a prompt will improve performance on a few isolated examples, but will result in lower overall performance on a more representative set of examples. Therefore, to ensure that a change has a net positive effect on performance, it may be necessary to define a complete test suite », summarizes OpenAI.

Important clarification, it is preferable to provide text external to the model as an example in order to better counteract its propensity to hallucinate. This phenomenon, which affects all generative AI, pushes them to generate completely false content while passing it off as truthful.

Source : OpenAI