Reverse prompting: how to get what you want from ChatGPT?
It’s not always easy to get what you want from ChatGPT. Yet, once again, it’s about making good use of our brains. In this article, I show you how, by reversing the established order, asking ChatGPT not to respond to our prompts, but to create them for us. A technique I use daily, which has transformed the way I solicit and utilize AI, whether it’s to refine the relevance of the responses obtained, innovate in learning, or navigate through the countless creative paths this technology offers.
What is reverse prompting?
“Reverse prompting” is a concept mainly encountered in the field of AI, particularly with language models and content generation systems like ChatGPT. This concept can be understood in different ways depending on the context, but generally, it refers to a method where the model is used creatively to generate inputs (prompts) from desired outputs, or to obtain information by asking questions in a way that “reverses” the usual approach of interaction with AI.
In the context of ChatGPT, instead of giving a prompt to the AI to generate a response, one could give it the type of response desired and ask the AI to generate the prompt that could lead to that response. This can be useful for understanding how to formulate prompts to achieve specific results, or for training the model to better understand and respond to complex queries by learning from its own prompt suggestions.
In practice, reverse prompting can be used for various applications, such as improving the accuracy of AI responses, creating games or educational exercises, or exploring new ways of interacting with AI technologies. The origin of reverse prompting can be seen as a response to two major challenges in the field of AI: firstly, the need to improve interaction between humans and AI models to obtain more accurate and useful responses, and secondly, the desire to explore and experiment with the models’ ability to generate creative ideas and solutions.
Principles of reverse prompting
Reverse prompting requires a real understanding of how AI models process information and generate their responses. The user must be able to think like the model, anticipating potential responses to different types of prompts. This method encourages experimentation and iteration, where different prompts are tested and refined until the desired result is achieved. To accomplish this, remember that ChatGPT does not answer questions but statistically predicts the continuation of what you’ve written in the prompt, much like when Microsoft Word suggests the end of the sentence you’re writing.
Exemples de prompting inversé dans différents contextes
- Education: A teacher looking to create exam questions that test students’ deep understanding of a subject could use reverse prompting to first generate the ideal answers or complex explanations, then work backward to formulate the questions that would lead to those responses.
- Marketing: A marketing team wanting to generate catchy slogans for a new product could, by first defining the emotions or key messages they want to evoke, use reverse prompting to identify the types of prompts that would lead to creative slogans matching those goals.
- Product development: Product designers can consider the ideal feedback or product features desired by end users. By applying reverse prompting, they formulate questions that could help them explore design paths that meet those expectations.
The 7 steps of reverse prompting
Reverse prompting is a form of ‘reverse-engineering’, where you start from the result to find the best possible starting point, imagining the steps in an inverse order!
This process is quite intuitive once you’re used to it, but to start from a generic base, here’s a detailed explanation of each step of the process:
1) Identification of the desired outcome
This first step involves defining precisely and clearly the type of response or result you wish to obtain from the AI. It’s crucial to be as specific as possible to guide the prompting process correctly. For example, if you’re looking for a creative solution to a problem, the desired outcome could be a list of innovative ideas.
- Difficulty level: 2 — Requires a clear understanding of the goal, but relatively direct and intuitive to start.
- TT tip: Be as specific as possible in describing the desired outcome. Use precise adjectives and clear objectives to guide your vision.
2) Designing the initial prompt
From the identified desired outcome, you formulate an initial prompt. This prompt is essentially a question or instruction designed to guide the AI towards producing the desired result. The quality of the prompt is determinant for the quality of the AI’s response. For example, to generate innovative ideas, the prompt could encourage the AI to think unconventionally or to apply concepts from one domain to another.
- Difficulty level: 3 — Requires some creativity and understanding of how the AI might interpret the instructions.
- TT tip: Incorporate relevant keywords related to the desired outcome in your prompt. Think about how you would pose the question to a human expert.
3) Testing the prompt with AI
Once the initial prompt is designed, it is submitted to the AI, which generates a response based on the prompt. This step is critical to assess the effectiveness of the prompt and the relevance of the AI’s response.
- Difficulty level: 1 — Technically simple process but requires careful examination of the results.
- TT tip: Keep an open mind and be prepared to be surprised. The first responses may not be perfect but offer unexpected perspectives.
4) Evaluating the answers
Several possibilities exist here: if the AI’s response matches the desired outcome, the process can be considered successful. This means that the prompt was well formulated and that the AI correctly interpreted the instructions. But if the response obtained is not satisfactory or does not match the desired outcome, it’s necessary to proceed to the next step to adjust the prompt.
- Difficulty Level: 3 — Qualitative evaluation of responses can be subjective and requires real discernment.
- TT Tip: Use an evaluation grid with clear criteria (relevance, creativity, specificity) to judge the responses. This helps to remain objective and focused!
5) Adjusting the prompt
This step involves modifying the initial prompt considering the responses previously obtained. The goal is to refine the prompt to get closer to the desired result. This might involve making the prompt more specific, adding details, or clarifying the instruction.
- Difficulty level: 4 — Requires a nuanced understanding of how adjustments will affect the AI’s responses.
- TT tip: Integrate feedback from the evaluation step to refine your prompt. Consider adjusting the level of detail, rephrasing, or adding clarifications.
6) Iteration
The process of adjusting the prompt and testing with AI is repeated as many times as necessary. Each iteration cycle aims to refine the prompt further until a satisfactory response from the AI is obtained.
- Difficulty level: 5 — Can be the most challenging due to the need for iterations and fine adjustments.
- TT tip: Be patient and persevering. Success often lies in repetition and gradual refinement of prompts. Each iteration is an opportunity.
7) Analysis of the results
Once a satisfactory response is obtained, it’s important to analyze the results of the process, including the observed benefits and limitations. This analysis can reveal insights into how the AI interprets prompts and generates responses, as well as the prompting strategies that work best to achieve the desired results.
- Difficulty level: 2 — Less technical but requires an ability for analysis and critical thinking.
- TT tip: Take a step back to assess the results as a whole. Reflect on what you’ve learned about AI, the process, and the subject matter.
This process shows in a generic, yet logical manner, the interactive and iterative nature of collaboration with ChatGPT and generative AI more broadly. Success depends on the ability to formulate effective prompts, interpret the AI’s responses, and adjust strategies accordingly.
Advantages and disadvantages of reverse prompting
Like all techniques, reverse prompting isn’t magic, and it’s necessary to use it with an understanding of its strengths and weaknesses.
Advantages
- Increased creativity: Reverse prompting pushes users and AI designers to think outside traditional frameworks, fostering a more creative approach to problem-solving and idea generation.
- Better understanding of AI models: Experimenting with reverse prompting, users learn to better understand how AI models process information and react to different types of prompts. This can improve the efficiency of future interactions with AI.
- Personalization of answers: This method allows for a finer personalization of responses generated by AI, as it starts from a specific desired outcome and seeks the most direct path to achieve it.
Limitations
- Complexity and time: The process of determining the ideal prompt can be complex and time-consuming, requiring multiple iterations to refine the prompt that produces the desired result.
- Requires in-depth knowledge of AI: To be effectively implemented, this approach demands a good understanding of the inner workings of AI models, which can be a barrier for less experienced users.
- Unpredictability of results: AI models can sometimes generate surprising or unexpected responses, even with carefully designed prompting. This can lead to inconsistent results or require additional adjustments.
The effectiveness of reverse prompting depends on the user’s ability to navigate the trees of possibilities, requiring a combination of technical skill, critical thinking, creativity, and patience.
Practical application: step-by-step example of reverse prompting
To illustrate the concept of reverse prompting in the context of obtaining a job offer for an HR innovation intern, let’s imagine we want to discover what type of prompt could prompt an AI model to generate such an offer. Instead of directly formulating a request, we’ll start from the desired objective (the job offer) and work backward to determine a prompt that could lead to this output.
Desired outcome
A detailed job offer for an HR innovation intern position, including responsibilities, required skills, and details about the company.
Reverse prompting approach
Ask the model to generate the question or series of directives that, in its opinion, would produce a job offer for an HR innovation intern as a response.
Example of reverse prompt
“Imagine you’re an HR expert looking to write an attractive job offer for an intern passionate about innovation in the HR field. What questions would you ask an AI model to create this offer, emphasizing innovation, technology, and sustainable development in the HR field?”
Expected outcome
The model could suggest a series of points or specific questions that would help identify the key elements to include in a job offer. For example, it might propose asking for details on ongoing HR innovation projects, the technological and interpersonal skills sought, or how the position will contribute to the company’s sustainable development mission.
Iteration
As always with generative AI, it’s important to consider that the provided response is a basis for further exploration and deepening. Even with reverse prompting, there’s no escaping this. It could involve asking for an improvement of the proposed prompt, an enrichment of the questions to be asked, or a challenge to the final result. And to tell you the truth, it works better if we do it for these three steps! The iterations may seem laborious and time-consuming, but remember, it’s the AI that’s working, you only have to think properly for a few seconds and a few steps, so it’s worth it.
Using this reverse approach, we can better understand how to structure our queries to obtain specific results from AI models, while exploring new ways to engage in conversation with these technologies. This can be particularly useful in innovative fields like HR, where the ability to ask the right questions can unlock valuable insights and encourage creative thinking.
Conclusion
By positioning ourselves not as requesters but as creators of prompts, we transcend the usual limits of AI to become active partners in the creation and innovation process. This approach allows us to rethink our interactions with technology, focusing on co-creation and experimentation, and opening ourselves to previously unimagined possibilities.
Far from being a mere exercise in style, reverse prompting invites us to a complex intellectual journey where our curiosity, creativity, and ability to anticipate AI responses become the keys to an enriched dialogue. It reflects the evolution of our relationship with technology, where the user is no longer a mere spectator but an actor in artificial intelligence, capable of guiding, questioning, and exploiting its potential in unprecedented ways.
[Article written and illustrated on February 25, 2024, by Jeremy Lamri with the support of the Open AI GPT-4 algorithm]
— —
If you are interested in the combination of web 3 with HR, and more generally in the societal challenges related to the emergence of the metaverse, AI, blockchain, quantum computing, and other emerging technologies, I invite you to subscribe to the dedicated newsletter that I hold on the subject, and to read the articles that I write:
