The Art of Thinking Well with AI – A Brief Practical Philosophy of Prompt Engineering

With the development of generative AI and the launch of GPT-3 in 2020, followed above all by ChatGPT at the end of 2022, a new way of interacting with information systems and computer terminals emerged: "prompting." For the first time in all of human history, it is now possible to converse fluidly with a machine, to query it in natural language on virtually any subject, or to ask it to generate texts almost instantaneously, and of a quality that most of us would be hard pressed to match. It is indeed perhaps impossible to overstate how much Large Language Models (LLMs) represent, as much as anything else, a revolution in interfaces, while simultaneously transforming the way we generate content.

In this context, a prompt is to a language model something like what a query is to a search engine, with the crucial difference that it goes so far as to radically change the very way we had grown accustomed to searching for information since the early days of the web — and indeed, with the democratisation of AI agents, the way a person ordinarily interacts with software applications. These agents take full advantage of recent advances in generative AI, whose uses they extend in a sense, since it is no longer merely a matter of generating content, but of performing tasks autonomously and on behalf of an individual or an organisation.

Thus, no human-machine interface has ever seemed so transparent. It now suffices for us to speak in order to instruct software applications to act on our behalf. But this illusion of transparency is at the same time an illusion of mastery. Do we really need to learn to talk to a computer, when those same computers have converted to what seems most natural in us — conversation? It is almost as if the balance of power had suddenly been reversed. Whereas in the past we had to assimilate the operating modes of an ever-growing number of software packages and adjust our working practices accordingly, the opposite is now happening. So much so that we have the impression that it is enough to type or say a few words to the machine almost intuitively to obtain whatever we desire. Yet what lies behind this illusion of mastery is in reality a shift towards a different kind of competence, rather than the disappearance of skills inherited from older interfaces.

The question is not how to use these tools, but rather how to use them well — in such a way that we obtain from them the best of what they are capable of producing, and above all, that we are in a position to critically assess what they produce and what we are led to disseminate at greater or lesser scale. To be at once more effective, but also lucid and responsible. For what is at stake is not only our employability and our adaptation to new working conditions, but above all our autonomy and our capacity to continue thinking for ourselves — against the temptation, out of laziness, convenience, or abdication, to delegate everything to the machine.

LLM, Generative AI and NLP. – As a first approximation, an LLM, or Large Language Model, is a type of artificial intelligence, falling within the category of generative AI, whose primary function is to help people produce — to generate — textual content automatically in response to an instruction called a "prompt." This prompt, unlike the keyword-based queries entered into a conventional search engine, is formulated in natural language. While LLMs belong to the broad family of so-called "generative" AI systems, they constitute only a subset of it, principally dedicated to text generation and the understanding of natural language, but also to those other types of formal and artificial languages that are programming languages, such as Python, Java, or JavaScript. Alongside LLMs there exists another category of generative models specialised in understanding and generating other types of data, such as images, videos, sounds, and music.

These generative AI models — the LLMs, which should no longer be confused with other types of generative AI models for images, sounds, or videos — have been trained on immense quantities of texts and content (training datasets), drawn predominantly from the web, but also, where necessary, from so-called "synthetic" data generated artificially by algorithms. It is precisely during this training phase that they construct, in a sense, a mathematical representation of language, and thereby become capable of handling a whole set of tasks generally grouped under the more generic term of Natural Language Processing (NLP) — such as understanding a text, producing a summary or a detailed explanation, performing translations, generating original content, or answering questions on virtually every domain of human knowledge. They have very rapidly evolved into multimodal systems capable of processing and combining several types of data, enriching interactions while also, and above all, extending their possibilities to an ever wider range of use cases in the fields of healthcare, education, entertainment, artistic creation, management, customer service, and logistics.

As the number of parameters increases — that is, the number of numerical values or weights that encode the syntactic and semantic aspects of language — these models develop increasingly remarkable emergent capabilities: solving complex mathematical problems, performing comparative and thematic analyses of ever larger text corpora thanks to extended context windows (the quantity of text the model can take into account simultaneously when generating responses), and even the emergence of statistical metacognitive capabilities. These allow them a form of self-criticism regarding the relevance of the content they generate, as well as the implementation of iterative processes aimed at improving their outputs — which raises the central question of the degree of genuine understanding possessed by language models.

Although this point is the subject of an increasingly lively scientific and philosophical debate among specialists, a long-held position, now contested, holds that a language model possesses no semantic understanding of the symbols it manipulates in the sense that a human being would understand the meaning of words. In other words, it relies primarily on statistical and probabilistic techniques in order to determine the probability of a sequence of words in a sentence. In short, an LLM has at all times only one real objective: to predict the next word or words in a text in such a way that the result is as credible as possible from the standpoint of its syntactic and semantic construction.

Thus, if we type the sentence "When the cat's away, the mice will…", the LLM will simply assign to each word in the vocabulary acquired during its training phase a statistical probability of occurrence in the given context, and select, for example, the word "play" as the most appropriate continuation of the proposed sequence. But what the model actually calculates in the background is the probability of millions of possible outcomes. In this case, the model proposes the verb "play" because the probability associated with that word is higher than that of other candidate words. It is precisely this capacity to combine terms with one another in an almost infinite variety of ways, according to a syntactically and grammatically coherent order, that makes language models excel at tasks such as simulating a conversation, performing logical reasoning or deductions, correcting, summarising or analysing texts, explaining a subject and answering questions — though always within certain limits.

It does indeed happen that the model proposes a syntactically malformed or semantically incoherent sequence of words, or even produces entirely false answers, while giving the impression of perfect command of its subject. This is what has been called, since the deployment of ChatGPT at the end of 2022, using a term borrowed from psychiatric vocabulary, "hallucinations." These are by no means a simple software bug, but an intrinsic property of the system's functioning. Since the model does not possess a structured and verifiable factual memory, it generates statistically probable word sequences rather than confirmed facts, which can produce false assertions formulated with apparent confidence. While the most recent models have considerably reduced this phenomenon — notably through the integration of web browsing modules and RAG systems — hallucinations remain a structural reality that an informed user must take into account, in full awareness of the use they are making of the model, the context, and the way they are querying it.

This capacity for processing and generation rests on a decisive architectural rupture, whose founding act is a paper published in 2017 by researchers at Google, entitled Attention Is All You Need. By introducing Transformers, a then radically new neural network architecture, this paper provided LLMs with their essential cognitive infrastructure: the capacity to capture mathematically the meaning of words and sentences by focusing on the relevant elements of a context, and by taking into account the position of each word in increasingly long sequences. It is this architecture that opened the way to the revolution in language models, and beyond that, to generative artificial intelligence, thanks to the capacity of these systems to produce content at scale and of a quality previously unmatched. The "T" in the acronym GPT refers precisely to this architecture.

The terms prompt and prompt engineering. – The word "prompt" derives from the Latin promptus, meaning "ready" or "prepared," while the Latin verb promere means "to bring forth" or "to produce." From this Latin root come the English terms for prompt, used both as a noun — "incitement," "cue," "invitation" — and as a verb — "to incite," "to encourage" — as in the sentence: "A mother must prompt her children to tell the truth." In computing, the prompt referred, in the 1960s and 1970s, to the line of text displayed by an operating system with a text-based interface, inviting a user to enter a command — such as deleting a file or creating a new folder — at a time when graphical interfaces had not yet become widespread. Computing was then the preserve of specialists — or "nerds," as they were called — who had to master a cryptic formal language in order to activate even rudimentary functions. And it was precisely against this elitist usage, and in a desire to democratise the use of the microcomputers that were beginning to appear at the time, that interfaces based on windows, icons, drop-down menus, and the mouse were popularised in the second half of the 1980s.

In the context of artificial intelligence, and in particular generative AI, a prompt is a text or a question inviting or requesting a machine learning model to produce a response. After the iPhone and touch interfaces, AI inaugurates, through natural language conversational interfaces, a radically new way of interacting with computers, of which the prompt is the new password. While it can vary significantly depending on the context or the task one is asking a model to carry out, it must nonetheless comply with a certain number of structural rules, whose role is to influence the nature and quality of the generated content so that it is as well suited as possible to a given user's expectations and situation. To put it differently, the quality of the response is directly determined by the quality and precision of the prompt.

In this context, prompt engineering will be defined as the art of designing and formulating instructions — prompts — for AI models, so that the latter produce specific outputs (texts, images, videos, music, multimodal content) that are as perfectly aligned as possible with the user's intentions. In other words, prompt engineering is the methodical and structured practice of interacting with AI systems in an optimised way, with a precise intended outcome and a command of the underlying mechanisms, in order to make full use of their capabilities. This requires first acquiring a deep understanding of how generative AI models work, such as OpenAI's GPT, Google's Gemini, or Anthropic's Claude. How can one make the best use of a tool if one does not know the reasons and the uses for which it was created? And how can one grasp the limits inherent in any tool — limits that are precisely defined by the uses for which it was designed — if one does not at the same time know how it functions? This is what Anne-Laure Enjolras, CEO of AGI Studio and expert in generative artificial intelligence, reminds us of in an article on how LLMs work, published in the magazine L'IA pour tous, in the December 2025 – January 2026 issue:

"This understanding is essential for any person or organisation wishing to use generative AI seriously: it enables one to improve prompts, to set expectations for quality, to anticipate biases, and above all to extract value without falling into naive fascination. It also makes it possible to understand why apparently minor adjustments — the structure of a prompt, for example — can sometimes radically transform the quality of results."

It is therefore unsurprising to see emerging, alongside the new professions of AI — data analyst, data manager, data miner, data steward, data designer, AI developer, machine learning engineer, AI architect — the profession of Prompt Engineer, which is attracting considerable interest with the proliferation of large language models for text generation (GPT, Claude, LLaMA, Gemini, etc.), image generation models (DALL-E, Midjourney, Stable Diffusion), audio generation models (Murf, ElevenLabs), and video generation tools (Runway ML, Pika Labs). Beyond the ability to understand the objectives of each project and to translate user needs into clear and concise prompts, as well as mastery of the capabilities and limitations of the models available on the market, all job postings insist, without exception, on the acquisition of specialist skills in artificial intelligence, whether technical, ethical, or legal. As if the proper use of generative AI tools — as Anne-Laure Enjolras reminds us above — necessarily entailed a prior understanding of how they function.

Some go so far as to invoke the notion of cognitive anticipation capacity to designate the aptitude a human being would have to model the way an AI is likely to interpret instructions and generate certain behaviours in response. To this are added broader competencies: critical thinking, text analysis, writing skills implying a certain rigour in the formulation of the prompt, the ability to take into account potential biases in the model, creativity, curiosity, a taste for experimentation, and finally a sufficient general culture to guide and evaluate the AI's responses on the subjects submitted to it — even though the AI itself can help us to deepen and discover them.

Fundamental principles of prompt engineering. – On the basis of scientific studies devoted to best practices in prompt engineering, the authors of the excellent short volume Apprendre à guider les IA, Mikaël Cabon and Cyril de Sousa Cardoso, recommend eleven key principles for becoming an expert in prompt engineering. These principles, grounded in the most recent research, can also serve as a foundation for developing an applied ethics of how we should make use of generative AI technologies. As such, they could serve as the basis for an ethical charter guiding the practices and uses of this type of AI. Beyond the pursuit of efficiency, they above all remind us that we, as users, bear a responsibility for the impact of what these tools can produce — not only at the individual level, but also collectively: every AI-generated piece of content that is disseminated contributes, in its own way, to shaping the informational environment in which we live, to forming our judgement, and to ensuring as far as possible the soundness of the decisions we make for ourselves and as citizens.

How indeed can we ensure that we use them in accordance with what they are and the way in which they were designed? How can we correctly interpret the apparent certainty with which they deliver the information we request? Are we truly aware of the potential negative effects they can have if we use them blindly, without taking into account the limits they contain, or without projecting onto them human characteristics they merely simulate without ever truly possessing? What is our direct share of responsibility for what these tools produce and for the quality of the content we communicate? In other words, to what extent can we trust them? All these questions insist on the need to interrogate which tasks we can delegate to them and which fall imperatively to us. They also invite us to define and respect a set of fundamental ethical values capable of guiding a use of these tools that is genuinely beneficial to all.

PRINCIPLE 1: Understand the model being used. This is the starting point of all learning in prompt engineering: a thorough understanding of the capabilities, but also the limitations, of any AI model. This means not only understanding the general operating principles of any generative AI model, but also the specific characteristics of the model one is using. It is important in this regard always to bear in mind that a generative AI model is not first and foremost a search engine, but a system capable of generating original content, which addresses a multitude of domains in the manner of an expert, and whose responses are produced by probabilistic computation rather than by retrieval of facts from a structured database.

PRINCIPLE 2: Prioritise specificity and clarity. Writing a prompt is in a sense equivalent to soliciting a response from an AI model, from which one expects a reply that conforms to one's expectations. But how can one obtain the desired response if those expectations are not clearly and precisely expressed? An AI is like a person in this respect: the quality of the response is proportional to the quality of the request, such that poor responses stem less from the AI itself than from the person querying it. One should not ask: "Write me a course on René Descartes," but rather: "Write me a course on the four rules of method as expressed in the Discourse on the Method, in a style suited to a secondary school student revising for the philosophy baccalaureate, and in 400 words."

PRINCIPLE 3: Incorporate details and examples. The term "task" will be used to refer to what you are asking an AI to carry out on your behalf. As already noted, it is crucial that this task be described clearly and precisely, specifying where possible contextual elements such as the role, the style, or the length of the response. A good way of clearly expressing what one wants is to provide the model with examples. This technique, which specialists in prompt engineering call "few-shot prompting," consists in illustrating the type of response expected, so that the model can infer the desired response pattern.

PRINCIPLE 4: Define the audience and the channel. It is equally important to specify the audience for the content and the channel that will be used to disseminate it. Is it a blog post? A course sheet? A revision aid? Is it addressed to young teenagers? To secondary school students? To specialists in a given field? These details directly condition the register, the degree of technicality, and the structure of the response produced by the model. In this context, one may draw on the notion of "persona," widely used by designers, developers, and marketing specialists. This refers to a fictional representation of a typical user or customer, intended to guide the design of a product, an interface, or a service.

PRINCIPLE 5: Use a conversational tone. One should avoid jargon, as well as familiar language or slang, with a primary concern: to express as clearly as possible, and in standard terms, what one is asking of the machine. A prompt well formulated in everyday, precise language will always be more effective than a technically sophisticated but ambiguous instruction.

PRINCIPLE 6: Ask open-ended questions. One should avoid using closed questions that call for a short, limited response of the "yes" or "no" type, thereby restricting the possibility of obtaining rich, detailed, and nuanced information. Open-ended questions, by contrast, allow for precisely this, and may begin with formulations such as: "Why?"; "How?"; "What are?"; "Describe to me?"; or "In what way?" But these same questions, which can force the model to surface ideas one might not have been able to anticipate oneself, can also be sources of errors, vague statements, or hallucinations, if they are not framed by a clear context, a defined posture, and an assigned role — all of which serve to steer the model's responses in a direction consistent with one's intention and to limit potential drift. From what point of view, or in the name of what identity, do you wish your AI to respond? In what context does it operate? What level of expertise can it claim?

PRINCIPLE 7: Employ iteration and feedback. Prompt engineering is an iterative process. It involves progressively adjusting one's requests in response to the replies provided, until the result is as close as possible to what one was seeking. Each exchange with the model is an opportunity to refine the formulation, correct an ambiguity, or enrich the context provided. This is a "test and learn" philosophy that values a spirit of experimentation, the capacity to critically question what one has produced at first attempt, but also the ability to halt the iterative process and judge that the result obtained, if not entirely consistent with the initial intentions, is in any case satisfactory enough to bring the work to a close. The lesson drawn from each iteration is also a means of deepening one's knowledge of the model being used, while the variation of formulations should encourage us to refine our command of language.

PRINCIPLE 8: Guide the output. It is possible to steer responses by writing the first word or first sentence of the expected output. One can also have recourse to what is known as "decomposed prompting," which consists in dividing a complex task into several successive sub-tasks, each of which is the subject of a specific prompt. If one wishes, for example, to write an article on a specific topic, one might first ask the model to provide a detailed outline of the article, then use each section of the outline as input to generate the complete content of each part, and finally bring together all the sections in order to refine the style and verify the overall coherence of the text.

PRINCIPLE 9: Experiment with different models. Since each model has its own specificities, one should test the same prompt on several models in order to see which produces the most satisfactory result. This comparison is particularly useful for high-stakes tasks, where the choice of model can make a significant difference to the quality of the output. It will thus be possible to compare models against one another across different use cases and form a precise idea of where each excels or shows its limitations. This experimentation should remain regular and ongoing in view of the rapid progress of models, which can change significantly several times in a single year.

PRINCIPLE 10: Learn through practice. Prompt engineering should be regarded as an art and a practice whose theory one continually refines over time, and in which everyone can benefit from the good practices of others. One can thus keep a journal of prompts that have yielded the most successful results, and gradually build up a personal repertoire of best practices that one has experienced first-hand. Without forgetting, of course, that models are constantly evolving, and that with them, practices may need to be adjusted as well. Experimenting also means testing one's practices in varied contexts and with different tools.

PRINCIPLE 11: Interpret and use responses. Finally, one should not take the responses provided at face value, but assess them with a critical eye by cross-referencing them with other sources of information. The responses provided are less established knowledge than raw information that one can then evaluate, reprocess, and use within a clearly identified framework over which one retains control. Prompt engineering does not exempt one from human judgement: it presupposes it.

The canonical model of a good prompt. – At the present time, there is no repository cataloguing all types of prompts. One can, however, draw on the most commonly cited categories, as we shall see below. There is, on the other hand, a consensus around the existence of a canonical model of an effective and universal prompt. Following this model rigorously makes it possible, in theory, to get the most out of generative AI systems, whether ChatGPT, Claude, Gemini, DeepSeek, Grok, or Mistral. These models are built on similar architectures and largely standardised technologies. Even if their performance sometimes varies significantly depending on context and use case, the way of interacting with them remains broadly the same — which facilitates comparison between these tools and the transition from one to another without having to relearn how to use them each time.

While this model, now widely adopted, can serve as a starting point and reference, it is nonetheless flexible. In other words, it is not a rigid form to be applied to every task, regardless of the level of complexity of the responses one expects. It is precisely these two criteria — the nature of the task and the degree of sophistication required — that should guide the way a prompt is constructed. Hence the importance of clearly defining from the outset what one expects from an LLM, and of adjusting the prompt accordingly — failing which one risks leaving it to the machine to fill, on the basis of purely statistical criteria, the grey areas that one's instruction has not managed to dispel.

In this context, the most rudimentary form of prompt — its degree zero, to borrow a term from linguistics applied to a literary text stripped of all stylistic convention — is what is known as "zero-shot prompting." It can be compared to the use we make, or rather used to make, of a conventional search engine such as Google or Microsoft Bing until not so long ago. In other words, it is a prompt stripped of artifice, consisting in posing a question or giving an instruction without providing any contextual element, any further specification beyond simply and directly asking the model for what one wants. Zero-shot prompting proceeds from the assumption that the question being asked is so elementary, and that the model has ingested such a wealth of knowledge and varied situations, that it does not need to be guided in order to respond to one's request. This is almost the exact opposite of the canonical prompt we are about to describe in detail. Does one really need to specify the context and the standpoint from which one wants an explanation of the following sentence: "The French state is forecasting a public finance deficit of 5% this year" — or indeed: "Translate this sentence into English"?

Since no context is specified, it goes without saying that zero-shot prompting must be able to draw on the model's capacity to generalise by reasoning, in a sense, along the lines of "all else being equal." In the case of the example mentioned, one could almost remove the word "French," or replace it with any other country, since what matters here is to understand what it means to "forecast a public finance deficit of 5%." The model will take the freedom, on the basis of purely statistical criteria, to choose not only the form of the explanation and its degree of technicality, but also its degree of depth. This also means that a simple reformulation of the request around the same task can not only change, but above all significantly enrich the model's response, by restricting its freedom of action and the field of possibilities within which it will have to formulate a reply.

Other, less rudimentary forms of prompt have thus emerged, which serve to specify the task — either by defining the role the model is to play and the output format of the expected response (the RTF model, for Role, Task, and Format), or by replacing the role with a situational framework (the CTF model, for Context, Task, and Format), thereby substituting for the question "What is the AI's identity?" the question "What situation is it in?" It is indeed not the same thing to have an AI take on the role of an economics professor specialising in public finance in order to obtain an explanation of the sentence "The French state is forecasting a public finance deficit of 5% this year," as to replace that role with a situational framework specifying that one wishes to have the same explanation in the context of the Maastricht criteria. Wherever the form of the response matters as much as the content, the RTF model will prevail, since it is self-evident that registers of language — that is, the manner of formulating a statement, lexical and stylistic choices — are strongly dependent on the speakers and the roles they play in a given communicative situation. If one asks a language model, for example, to analyse the legal risks contained in a contract, one would be well advised to have it take on the role of a lawyer specialising in contract law. If, on the other hand, the aim is to obtain three pieces of advice for developing a rapid recruitment strategy in the context of a twenty-person SME growing at more than 20% per year, operating in the IT consulting sector in data science and artificial intelligence, the situational framework will here be decisive.

It will readily be understood that these two prompt models are not antithetical. Rather than opposing them, one should prefer to combine them, or even enrich them in accordance with the complexity of the task in question. Bearing in mind these few remarks, here are presented, in the appropriate order of relevance, the seven building blocks of a reference prompt, which ultimately do no more than reproduce the architecture of a request one would address to a colleague in the context of a task one was asking them to carry out on one's behalf.

Block 1 — The role: The first step consists in assigning the model a precise role directly aligned with the context of your request. This is, in a sense, the posture you are asking the model to embody in the context of a given task. Why is this point absolutely fundamental? Because it enables the model to activate a register of knowledge — that presumed to be associated with a given socio-professional category — as well as a specific tone, both of which are directly linked to the function you are assigning it. Who would you ask to write an article on the history of search engines, if not a computer scientist or a historian specialising in the history of the web? A typical formulation might thus begin as follows:

"You are a historian of technology, specialising in the history of the web, recognised by your peers for having written a definitive work on the history of search engines from 1945 to the present day."

By defining this role as precisely as possible, one thereby circumscribes the model's discursive space, and ensures that it will mobilise knowledge entirely suited to the function one is asking it to assume.

Block 2 — The context: The second step, equally important and continuous with the first, consists in providing the model with the situational information it needs to understand the environment of your request. This is, in a sense, specifying to the model why you are asking it to behave as a historian, a marketing professional specialising in SEO (Search Engine Optimisation), or a web developer. Or, if you prefer: "To what end?" The question you must answer is the following: "For what reasons are you asking the model to behave as a historian of technology specialising in the history of information retrieval?"

Here too, the analogy with a colleague seems obvious. It seems entirely natural to specify to a person the context and reasons for the task you are asking them to perform. A task is by its very nature always oriented towards an end. It has a precise objective and is situated within a given context — which is, moreover, perfectly implied by the notion of "agent," whether human or artificial. Why should it be any different for a virtual agent? The last thing one should do is assume that the model will guess the situation in which it is supposed to act. Keeping our example in mind, we might complete the role assigned above with the following formulation:

"I am a journalist in charge of the 'technology' section, writing for a broad audience not necessarily familiar with computing, and I wish to publish a series of articles devoted to the key applications that have shaped the history of the web since it became widely accessible to the general public in the early 1990s, up to the present day."

Block 3 — The task: It is only at the third step, having specified the role and the context, that one can at last enter the heart of the prompt. What is the task you are asking your virtual agent to carry out on your behalf? Prompt engineering can thus be conceived as a school of rigour that obliges one to formulate as clearly as possible the type of response expected from a generative AI — such that the quality of responses depends as much on the model as on the quality of the instructions provided. In other words, one shares responsibility for the relevance of one's model's responses.

On this occasion, one may borrow from the classical aesthetics of a Boileau one of the most fundamental rules of prompt engineering: "What is well conceived is clearly stated, and the words to say it come easily." In other words, the precision of expression and the definition of the task you assign to your AI can only arise from the clarity of your thought. This means that the task will be carried out all the more effectively the more perfectly the way it has been written and the choice of words used correspond to your intentions at the moment of drafting the prompt. One should therefore avoid vague and overly general formulations such as: "Tell me about search engines" or "Say something about Google." Ambiguous and imprecise formulations should also be avoided in favour of precision and coherence, while bearing in mind that these requirements remain relative to the role and context assigned to the AI.

This leads to two essential rules which may seem counter-intuitive to an uninitiated user: 1) in order to correctly define a field of action for an AI, one must be familiar with that field and possess basic knowledge of it; 2) it is this more or less thorough knowledge of the domain in question that makes it possible to choose the appropriate words to describe a task, most often in connection with a role and a context. If, for example, you wish to prepare a course introducing students to the philosophy of mind and current debates on the conditions for the possibility of artificial consciousness, it is imperative to possess both technical knowledge of how current AI systems work, and a solid grounding in philosophy. It is only on this basis that you will be able to formulate the task adequately and cast a critical eye over the response provided. In the context of our ongoing example, this might yield the following formulation:

"Your mission is to write a text on the history of search engines from the birth of the publicly accessible web in the early 1990s up to the advent of the first generative AI systems and the emergence of new forms of search such as Google's Search Generative Experience. While following the chronological order of events as closely as possible, you will above all emphasise the principal technological ruptures that marked this period, and explain how generative AI is fundamentally changing the search experience. You will conclude with a discussion of the now debated question of a supposed end of the web, and what precisely should be understood by this expression."

Block 4 — The constraints: This is the most frequently overlooked block, and yet one of the most decisive. A good prompt consists not only in indicating what we want the AI to do, but also what we do not want it to do. In other words, constraints are intended to frame what we are seeking at all costs to avoid, and contribute in a sense to clarifying and circumscribing the task by indicating what it must not be. They thus participate in the process of framing the work to be done, forcing the model to concentrate its efforts where we want them, and to avoid — to borrow a term familiar to aspiring philosophers — going off-topic, or, somewhat less seriously, engaging in pointless digressions that add very little value to the question posed and prevent it from being treated in depth. In cases where factual accuracy and verification are absolutely indispensable, constraints can also oblige the model to build its responses on the basis of reference websites or authoritative blogs, thereby avoiding any form of speculation or hallucination. They thus call upon the model to acknowledge certain of its limitations, by refraining from venturing into speculative hypotheses or unconfirmed information. To continue with our ongoing example, the following constraints might be added to the task:

"It is absolutely essential that the facts and events you recount are verified and sourced from reference sites relevant to the topic of search engines. You will ensure that your information sources are authoritative in the field in question, and that you cite them. With regard to the debate around the death of the web, which at this stage remains largely a matter of opinion, you will endeavour to provide as exhaustive an account as possible of the current debate, drawing on and citing a range of expert views. You will finally avoid technical jargon as far as possible, in order to address the widest possible audience of non-specialists."

Block 5 — The format: If the first four blocks — role, task, context, and constraints — are there to orient the conceptual content of the response that the generative AI must produce, the last two — format and tone — relate more to the manner in which that response should be formulated, presented, and ultimately structured. We encounter here the classical distinction between, on the one hand, substance — which belongs to the realm of meaning, content, and ideas — and, on the other hand, form — which belongs to the realm of style, syntax, and rhetoric. And what could be more natural, since this fundamental duality lies at the heart of the architecture of every form of expression and communication, whether narrative, descriptive, poetic, explanatory, or argumentative in character.

Making this essential distinction explicit is, moreover, far from trivial or ornamental in a text devoted to a practice whose purpose is above all operational and concrete, since it ultimately governs the quality of the response provided by the AI. In other words, both must be perfectly aligned, since the final form of the content generated by the model must ultimately reflect it. One would not require, on a given subject, the same type of response depending on whether one is addressing a community of researchers or an audience of non-specialists — or if the response one is seeking is intended merely as a source of information for another text one is about to write. Hence the need to specify as precisely as possible such essential presentational elements as the expected length of the response, its overall structure (titles, introduction and conclusion, chapters, paragraphs, bullet-pointed lists, summaries of key points to retain, etc.), as well as any other formal element that must be consistent with the role, context, and posture you are asking your AI to embody.

"Your text will begin with a title and a lead paragraph summarising the essentials. You will then proceed to developments comprising an introduction setting out the context, followed by paragraphs recounting facts and dates in chronological order, to which you will add explanations and analyses. You will endeavour as far as possible to define the technical terms you use, never forgetting that you are writing for an educated but not necessarily expert audience. Your conclusion will summarise the key points to be retained, and may open onto questions currently left in abeyance but which, according to experts, could prove decisive for the future of the web and search engines. The total text should be approximately 1,000 words."

Block 6 — The tone: While the format emphasises the material, structural, and concrete elements of a text in relation to the category to which it belongs — is it a blog or newspaper article, an academic essay, a short story, a novel, an email, an advertising brochure, a briefing note, a case commentary, a speech, a podcast, a product description, and so on — the tone expresses the tonality of the text proper, which may be neutral, explanatory, familiar, analytical, argumentative, dialectical, imaginative, inventive, metaphorical, humorous, or emotional. Each textual format, far from being unconstrained, has its own specific conventions comprising both purely formal elements, such as length and structure, and stylistic ones — that is, a particular manner, more or less personal, of conveying a message and making use of language. In this context, tone can more concretely be defined by the choice of words, the rhythm, and the emotional colouring one wishes to give the text. In other words, while format indicates what the reader sees, tone refers to what you want them to think and feel. Hence the essential questions: What is the intention of the text I am asking the AI to write? What effects am I seeking to produce on the reader? Am I trying to instruct, to amuse, to convince, to persuade, or to move? In the context of our article on search engines, we might formulate the tone as follows:

"Adopt an impersonal and journalistic tone. The aim is to inform, instruct, and explain by drawing as far as possible on incontestable objective facts that punctuate the history of search engines. Sentences should be clear, concise, precise, and easy to understand for an informed reader who is not familiar with the subject. Avoid technical vocabulary used by insiders, and favour everyday language."

Block 7 — The example or reference: The final block, optional but often recommended depending on the nature of the task, consists no longer in telling the AI what to do and how to do it, but in directly providing it with one or more examples of the final result expected from it. In other words, one provides the model with a concrete illustration of what it should produce — one that, in a sense, reinforces and anchors in something tangible what until that point had remained in the realm of indication, description, and explanation. It is a little like teaching a law student to produce a case summary by asking them to read an ideal case summary (the reference), rather than a theoretical explanation of what a case summary's structure should contain.

The example thus draws on another remarkable and essential capacity that generative AI models in general, and language models in particular, possess: the aptitude to imitate and reproduce human behaviours. This capacity, commonly called "mimicry," and which constitutes the very fundamental characteristic of this type of AI, allows them not only to write and converse in the manner of a human being, but also to produce content that imitates the structure and style of reference documents submitted for their analysis. This aptitude for imitating virtually any style — which literary scholars call "pastiche" — extends to the ability to write, not without some awkwardness, in the manner of any writer or literary school, or to engage in any stylistic exercise: academic dissertation, literary criticism, research paper, journalistic article, political speech, free verse, slam poetry, autobiographical narrative, and so on.

This capacity, now well known in AI, can constitute an excellent means of standardising the writing of texts belonging to the same type, with a view to ensuring that all content produced shares the same editorial identity across time and authors, regardless of the themes addressed. Provided, however, one does not forget that a language model is liable to reproduce not only the format and tone, but also less immediately apparent and more difficult to detect elements: the ideas — not to say the ideologies — the values, the biases, the cultural preferences, and the stereotypes that are inevitably inscribed in a text. In certain cases, the choice of one or more examples should preferably be accompanied by a critical reading aimed at identifying the presuppositions and opinions they contain without ever explicitly naming them. This is an essential characteristic of the critical thinking mode.

To conclude with the example of writing an article on the history of search engines, it will be possible to attach to our prompt documents written in the format and style we wish to see the model apply. Our prompt will be all the more coherent and effective the more exactly the example provided corresponds to the task, format, and tone previously described.

"Attached for reference is an article from the 'Technology' section of the newspaper that is not devoted to search engines, but which perfectly exemplifies the format, style, and editorial line I want you to follow."

From this presentation of the canonical prompt model in seven key building blocks, two further fundamental methodological rules can be derived — rules that serve as a reminder that a good prompt does not consist of a series of boxes to be ticked, or a form whose fields are to be filled in piecemeal and without regard for the overall logic. These two rules can be formulated as follows:

• Rule 1 — Separate to avoid confusion: In the interest of optimal readability on the AI's part, some experts strongly recommend visually distinguishing the various components of a prompt by explicitly naming each block and using delimiters or tags. One may thus use square brackets, angle brackets of the type <Constraints>, <Format>, or XML-style tags such as <tone>…</tone>. The idea is to use a consistent and uniform format to signal to the model where each section of the prompt begins and ends, allowing it to process each part without confusion or interference between them. This practice is all the more indispensable the longer and more complex the prompt being drafted. In other words, the greater the number of instructions, the higher the risk that the model will hesitate between two instructions located in different sections of the prompt that appear to contradict each other. An instruction relating to tone might, for example, contradict the role you have assigned to your model.

• Rule 2 — Think of the prompt as a coherent whole: The decomposition into seven blocks is above all a methodological and pedagogical tool. It must not obscure the fact that these seven parts constitute in reality one single prompt, each component of which must resonate with the others and mutually reinforce them. A reader coming to your prompt should immediately notice the overall logic connecting the way you describe each of the instructions you give the model, and should be able to effortlessly deduce the relationships between the role, the context, the task, and the constraints, as well as between the instructions relating to the form you wish to give to the final output (format and tone) and the elements more directly relating to the substance already mentioned. Is the role suited to the context? Do the constraints frame the task without contradicting it? Are the format and tone consistent with the role and my target audience? And so on.

Reading Descartes: still useful and relevant. – At bottom, the rules set out by René Descartes in the second part of his Discourse on the Method (1637), enunciated nearly four hundred years ago, could serve as meta-rules for the practice of prompt engineering. They are often reduced to four key terms, each of which encapsulates one of the rules: 1) the rule of evidence; 2) the rule of analysis; 3) the rule of synthesis; 4) the rule of enumeration. And how could it be otherwise, given that the techniques of prompt engineering ultimately do no more than translate into practice cognitive operations that all proceed from the same demand for rigour, clarity, and method in the processing of information.

"The first was never to accept anything as true that I did not know evidently to be so: that is, carefully to avoid precipitancy and prejudice, and to include nothing more in my judgments than what presented itself to my mind so clearly and distinctly that I had no occasion to doubt it. The second, to divide each of the difficulties I examined into as many parts as possible, and as might be required in order to resolve them best. The third, to conduct my thoughts in an orderly way, by beginning with the simplest objects and the most easily known, in order to ascend little by little, as it were by steps, to the knowledge of the most complex; and assuming an order even among those that do not naturally precede one another. And the last, everywhere to make enumerations so complete, and reviews so general, that I was assured of leaving nothing out."

While they cannot be taken, without a degree of anachronism, as a series of instructions for the design of a prompt, and provided they are reinterpreted in a context that is no longer that of Descartes's philosophy of knowledge — still steeped in a metaphysics and a conception of scientific method that are no longer ours today (one that marginalises experience and experimentation) — they can still serve, at the very least, as second-order rules — hence the use here of the prefix "meta" — or as fundamental principles for the practice of prompt engineering. After all, Descartes, greatly influenced by the rigour of mathematics, and in particular geometric analysis and algebra, was himself in search of a simple and universal method capable of guiding him in the production of knowledge.

The precepts bequeathed to us by the author of the Rules for the Direction of the Mind — a work composed around 1628 and published posthumously in 1701, which constitutes the first state of the method Descartes would formalise in the Discourse — should therefore be taken less in their original sense as a philosophical foundation for the modern science that was then emerging, and whose experimental dimension he, in a certain sense, underestimated, than as a restricted set of general principles to be brought to bear in any intellectual endeavour. Keeping in mind what precedes, and taking up each of the precepts of the Cartesian method as set out in the Discourse, it is possible to establish, in a summary fashion that would merit further development, the following transposition:

• The rule of evidence: The first precept, that of evidence, points to a twofold demand. On the one hand, a demand for clarity and distinctness in the formulation of the prompt itself: each instruction must be drafted so clearly and precisely that the model cannot interpret it erroneously or ambiguously. This is the direct application of the Cartesian requirement of clarity and distinction, but transposed no longer to ideas and judgements — as is the case in Descartes — but to the prompt. On the other hand, a demand for truth in the evaluation of results. One must never accept at face value what the model asserts with confidence without verifying its accuracy, and must ensure the soundness of the facts advanced. One may thus force the model, at the very moment of constructing the prompt, to check its sources and self-evaluate its response before delivering it.

• The rule of analysis: The second precept, that of analysis, refers directly to the decomposition of the prompt into clearly distinct, exhaustive, and precisely formulated parts. Let us recall that the etymology of the word "analysis," from the Greek analuein, meaning "to loosen," refers to the action of separating what was joined or bound together. The rule of analysis in Descartes indicates that the division into several key questions must be carried to a degree of granularity such that it permits exhaustive treatment of the subject. In the context of prompt engineering, it suggests that when faced with a complex subject, one should multiply the prompts into as many sub-themes as the subject may contain, rather than attempting to treat it all within a single instruction, which would risk generating not only confusion but also leaving to the model the statistical choice of filling what the prompt states confusedly. This would, in return, have the effect of diluting the precision of the results obtained.

• The rule of synthesis: The third precept, that of synthesis, which is the counterpart of the second — consisting in retaining an overview and then reassembling what was first separated — can be interpreted as a principle of methodical treatment consisting in approaching a subject first in its simplest and most accessible elements, then progressively ascending towards the most complex aspects, until the subject is treated exhaustively and according to an order of increasing complexity. There is thus a relation of dependence or deduction between the various elements, such that some can only be treated on condition that others have been treated beforehand. Applied to prompt engineering, and in the context of a subject involving some degree of complexity, one should endeavour to decompose it into simple and distinct elements (the rule of analysis), then build up a step-by-step understanding of it as comprehensively as possible, ensuring that each stage prepares and illuminates the next (the rule of synthesis).

• The rule of enumeration: The fourth and final precept, that of enumeration, is an invitation to conduct an exhaustive and systematic verification of the whole: to ensure that each instruction in the prompt has been properly respected by the model, that nothing has been omitted in the response, and that the outputs are complete and consistent with the initial intention. It is a check for completeness — a rigorous review of all the parts — from which there may follow, if necessary, an iterative process aimed at correcting any gaps or discrepancies observed between the initial intention and the result obtained.

We may therefore conclude this presentation of prompt engineering by noting that the four rules of the Cartesian method — the rule of evidence, the rule of analysis, the rule of synthesis, and the rule of enumeration — each find their correspondence in turn in several advanced prompting techniques: specifically, in order, "Chain-of-Verification prompting," "Self-reflection prompting," "Decomposed prompting," "Thread-of-Thought prompting," "Least-to-most prompting," and "Self-consistency." It should nonetheless be clarified that this correspondence is not strictly one-to-one: a single rule may illustrate several techniques, and a single technique may fall under several rules at once. I will now take up each of these prompting techniques and present them briefly, without going into detail or the possible variations of each, my aim being above all to highlight their connection to the four precepts of the method as set out in the second part of the Discourse on the Method. These techniques are not substitutes for or alternatives to the ideal structure we have previously set out, but rather ways of orienting it, refining it, or simply enriching it.

Chain-of-Verification prompting: This technique, referred to by the abbreviation CoVe, consists in asking the model to generate an initial response to a question posed, while implementing a protocol of self-verification of the facts or ideas advanced. It is structured around four key stages: 1) the completion of the requested task; 2) the generation of verification questions that make it possible to test the accuracy of the points made in the response; 3) the answering of the test questions; 4) the generation of a final response, corrected if necessary and more reliable. Particularly important in cases where factual accuracy is essential — notably in combating hallucinations — it can be regarded as the technical embodiment of the rule of evidence and of the concern for truth it is supposed to guarantee, even if in Descartes the criteria of truth are psychological in nature. We shall retain here only the Cartesian demand for truth, which consists in accepting nothing as true that has not been subjected to rigorous examination. Returning to the example of our ideal prompt on the history of search engines, one might add the following instruction to the model:

"Generate your own verification questions attesting that the facts recounted in the article have been checked and sourced, then adjust the final text accordingly to guarantee its accuracy."

Self-reflection prompting: In the same spirit, Self-reflection prompting is a simpler variant of Chain-of-Verification prompting, with the difference that the model generates a first response, then evaluates and critiques that response without having to generate test questions — hence the notion of "self-reflection" — before finally providing an improved and optimised final response. The self-evaluation process may bear on the accuracy of the content as much as on other, more formal criteria such as format, clarity, or tone. Here too, a parallel with the rule of evidence may be drawn, provided one does not anthropomorphise the internal self-evaluation process, which consists only in a set of algorithmic steps, involving no reflective consciousness, still less a phenomenal one. One might thus add to the end of the prompt an instruction along the following lines:

"After drafting the first version of the article, evaluate and critique that version by checking the clarity, tone, precision, and coherence of the format, then propose a final improved version."

Self-consistency prompting: A third technique, which can likewise be linked to Descartes's rule of evidence by virtue of the concern for truth it manifests, consists in verifying both the formal and the factual dimensions of the truth of a judgement, statement, or text. Philosophers generally distinguish indeed between material truth and formal truth. The former refers to the classical definition of truth, which has been defined notably since Thomas Aquinas (1225–1274) as the conformity or adequation between a proposition or statement and the reality it describes — the author of the Summa Theologica spoke of intellect or representation rather than of proposition, a distinction that changes nothing essential to the definition — while the latter no longer insists on the confrontation of the propositions of a given argument with reality, but on the manner in which those propositions are logically articulated with one another. Following the philosopher and logician Robert Blanché (1898–1975), one may distinguish the validity of an argument (formal truth) from the truth of the propositions that constitute it (material truth). A syllogism may thus be formally valid — its propositions follow from one another according to correct logical necessity — while resting on factually false premises, which renders its conclusion false despite the rigour of its form.

This distinction is essential, since it constitutes two important criteria of truth that must not be confused — what is materially true may be formally false, and vice versa — and above all because it is exactly this same distinction that underlies Self-consistency prompting. It consists indeed in asking the model to vary the same query multiple times and to identify which of the results is statistically the most probable. An example prompt might be:

"Tell me who is the author of the phrase 'Every truth is false as soon as one is satisfied with it.' Generate ten different ways of formulating the question, and return the answer that appears most frequently."

The Self-consistency technique is thus particularly suited to questions where factual truth, but also the formal validity of a text or argument, are especially critical. It takes advantage both of the limitations of a language model and of its probabilistic and statistical mode of generating content. Indeed, if one asks the same question ten times of a model, varying the formulation without changing the fundamental meaning, then the correct answer will statistically be the most frequent one. It is in a sense as though the truth and validity of an argument were submitted to a majority vote drawn directly from the corpus of texts on which the model was trained. Is it necessary to recall that a language model is not capable of distinguishing, as a human being would, the true from the false? That it has never been confronted with the facts it describes, except through the words it chains together probabilistically? The surest means is therefore to submit it to a factual base — via a RAG or a web query, for example — and to use its text-summarisation capabilities in order to obtain the most reliable response possible.

Decomposed prompting: As its name suggests, Decomposed prompting consists in taking a general and complex theme whose treatment requires decomposition into several sub-themes or elementary sub-tasks, each of which will be the subject of a separate, specific prompt, in a logically justified order of dependence. The results obtained will then be recombined to form a coherent whole. This is a technical transposition of the Cartesian rules of analysis and synthesis, which articulate themselves within a single dynamic comprising three successive thought processes: 1) decompose a whole into as many distinct elements as it is necessary to do in order to treat it exhaustively — this is the rule of analysis; 2) treat each of these elements according to an order of increasing complexity, ensuring that each stage prepares and illuminates the next — this is the rule of synthesis; 3) finally recompose the whole by ensuring that each part articulates coherently with the whole of which it forms a part — this again is the rule of synthesis in its recapitulative dimension.

Here one can especially see that the practice of Decomposed prompting implies, in a certain sense, having a fairly clear and comprehensive knowledge of the subject one wishes to treat. In the case of our article on the history of search engines, one would first need to define the major stages and key moments, then create a prompt for each of these stages, ensuring that the model follows a consistent editorial line and a similar tone throughout. It is precisely this knowledge of the history of search engines that enables one to formulate coherent prompts, and then to assemble the responses into a final, complete output consistent with the initial intentions.

This technique is potentially limitless, since it will be possible to subdivide this history with greater or lesser precision, or even to envisage sub-themes within each period, such as the evolution of algorithms, the arrival of advertising, business models, the impact of social networks, the evolution of SEO practices, the importance of privacy and the protection of personal data, the emergence of conversational search, and so on. All these sub-themes can be inserted into a chronology that recounts the history of the web and search engines, and can be the subject of specific and tailored prompts.

Thread-of-Thought prompting: As a variant of Decomposed prompting, it is also possible to associate the Thread-of-Thought prompting technique with the rules of analysis and synthesis, since it consists in encouraging the model to set out in detail the various stages leading to a given result, without ever losing sight of the overall logic. In the analysis of a document, for example, one would ask it to identify the themes and then the main thesis, before developing an explanation following chronological order and highlighting the main ideas chapter by chapter, before concluding with a synthesis of the key lessons. This enables the model to maintain a guiding thread between each part of the document and to preserve the coherence of the whole. It may also be noted that this technique has a secondary connection with the rule of enumeration, insofar as the guiding thread it imposes on the model is also a way of ensuring that nothing is omitted in the treatment of the subject.

In the context of our article on the history of search engines, one would indicate to the model precisely how one wishes it to construct the final text, by breaking down each of the stages it must follow, and above all by asking it to justify its responses. In other words, the model is forced to explicitly expose the logic it follows in elaborating its response, while simultaneously constructing the content. One might thus begin by asking the model to list the pioneer search engines before the arrival of Google and its PageRank, to justify the order in which it presents them, and then to explain why they played an important role in the history of the web. It might then move on to how Google built its hegemony in the 2000s–2010s up to the arrival of generative AI, and then explain and justify its responses. What is important to understand here is the dual mechanism of decomposition and justification at work, since the model is asked to follow precise steps without ever losing sight of the overall logic — the final result — while at the same time being invited to make explicit the reasons that led it to provide one or another result.

Least-to-most prompting: The Least-to-most prompting technique is perhaps the one that most faithfully recaptures Descartes's third rule of synthesis. It consists indeed in progressively deepening a problem or a given theme by starting with simple and general questions, then moving towards ever more precise ones, following a logical order of dependence, whereby the sub-themes identified in each of the model's responses call for further questions that gradually clarify the initial question. This technique ensures that the model follows the logical progression its user imposes upon it, working through each of the stages necessary to its resolution, until one judges that all the points relating to the initial question have been exhaustively clarified. Least-to-most prompting is therefore particularly suited to cases where one wishes to explore a question of which, at the outset, one has only a vague and uncertain idea.

Unlike Decomposed prompting and Thread-of-Thought prompting, which imply a certain prior knowledge of the subject, Least-to-most prompting is an exploratory technique that makes it possible precisely to distil a given theme, so as then to treat it using other techniques such as those described above. One might thus begin with a general question of the type: "When did the first search engines appear?" and then, drawing on the responses already provided by the model, progressively deepen the theme to its decisive point — which is that of the call into question of the existence of the web as it was envisaged in 1989 in the mind of its inventor, the British computer scientist Tim Berners-Lee, while he was working at CERN (the European Organisation for Nuclear Research).

It may be observed that the three techniques of Decomposed prompting, Thread-of-Thought prompting, and Least-to-most prompting all participate in the same process of progressive deepening, without however containing within themselves a final self-verification mechanism comparable to the rule of enumeration in Descartes. It is indeed important to insist on the logical and ordered character of the Cartesian method, each moment of which builds on the preceding one that makes it possible. The precept of enumeration should thus be regarded as a rule that in a sense closes and controls the process of elaborating knowledge, in order to ensure that all dimensions of a problem or a given theme have been treated exhaustively. This essential function can be fulfilled by a prompt I would call "reflexive," in the sense that it involves inviting the model to interrogate its own responses, and to evaluate by itself whether the initial subject has been treated in its entirety and exhaustively. We might submit the following open question to it:

"Read carefully through all the text produced. Carry out a complete enumeration of every aspect and dimension of the question treated. Review each part, each argument, each stage, and identify any points or perspectives that may not have been taken into account, making sure to omit nothing."

It is, however, important to keep in mind the limited character of this type of prompt, which is called "reflexive" here only by analogy and not by virtue of a cognitive process that would be in every respect similar to the human brain. The processing of the prompt, while inviting the model to return upon itself, or in any case upon what it has produced, does not induce any form of consciousness — of the form of consciousness that philosophers call precisely "reflexive" — and proceeds, like everything an LLM produces, only from an algorithmic and probabilistic form of self-evaluation.

Chain-of-Thought prompting: If Least-to-most prompting embodies the rule of synthesis in its progressive and constructive dimension — moving from the simple to the more complex — it is also possible to link to this same precept another prompting technique that explores another of its facets, no less important, which is that of deduction. We have indeed sufficiently underscored the influence that mathematics played in the elaboration of the rules of the method, which is clearly apparent in the passage of the Discourse that follows the presentation of the four rules, offering, in a sense, a condensed commentary on them and making explicit the entirely deductive logic that animates them:

"Those long chains of reasoning, quite simple and easy, which geometers are accustomed to using in order to arrive at their most difficult demonstrations, had given me occasion to suppose that all the things that can fall within human knowledge follow one another in the same way; and that, provided only that one refrains from accepting any for true that is not true, and that one always observes the order required to deduce them from one another, there can be none so remote that one will not eventually reach it, nor so hidden that one will not discover it."

It is exactly this logic that one finds at work in the Chain-of-Thought prompting technique, which moreover takes up very exactly Descartes's expression of "chains of reasoning," since it involves inciting the model to detail all the steps leading to a given result, and to exhibit the chain of logical operations carried out in the resolution of a problem. It is therefore particularly suited to the treatment of problems of logic or mathematics. One might thus, for example, ask the model to solve the following problem: "Five sevenths of a cake are divided into ten equal portions. What fraction of the whole cake does one portion represent?" — asking it to provide all the intermediate steps that structured its reasoning. Or, in an educational context, to solve an equation of the type 2x + 3 = 11, and to ask it to explain each of the operations performed to prove the equality of the two terms.

Does this mean that the use of the Chain-of-Thought technique is only valid in cases where the expected result is logically constrained and unique, as is the case for problems of logical reasoning or the solving of mathematical problems? In other words, is Chain-of-Thought only applicable if each step of an argument follows necessarily from a preceding one, as is the case with a correctly formed syllogism? It would thus admit only one type of argumentation, and would at the same time reduce all the problems we encounter in our professional and private activities to problems of formal logic — which is naturally not the case, and was not at all Descartes's intention, who took mathematics more as a school of rigour than as a kind of master key for treating all the world's problems.

Yet Chain-of-Thought can also be used in argumentative registers other than those of logic or mathematics, with the aim of exhibiting the logic underlying the model and the precise reasoning it followed in arriving at a solution. One might thus describe a company's situation to a model and ask it what might be the best strategic option to follow, listing exhaustively all possible options, examining the strengths and weaknesses of each, and justifying its final response. In the context of a classic philosophical problem such as the question of happiness — which cannot simply be reduced to an exercise in formal logic — one might ask the model to set out a step-by-step argument, starting from a common definition of happiness (the doxa, that is, common opinion), then examining and confronting each of the major theories one by one, before concluding in an argued fashion.

Prompt engineering is a discipline that is neither fixed nor settled, and which is expanding rapidly, continuing to evolve alongside the models in the years to come. In barely less than five years, researchers and practitioners have developed a genuine toolkit of varied techniques, each designed to solve a problem or address specific challenges whose only limits are ultimately curiosity, imagination, and experimentation. Far from encouraging the delegation and abdication of our cognitive faculties — which must be preserved, even protected at all costs — language models can also serve as formidable means of stimulating reflection, deepening subjects, making discoveries, and investigating new ideas we could not even have conceived of not long ago. Provided one does not ignore the fundamentally ambivalent character that the use of any technique contains — both remedy and poison (see my article The Techno-Utopian Project of Dario Amodei Tested Against Jacques Ellul's Philosophy of Technology, and in particular the essential notion of the "Pharmakon") — and that one learns to use them in such a way that they stimulate reflection, rather than paralyse it.

As a technique of self-expression in the same vein as writing, painting, photography, or cinema, linguistic models do not in any way constitute systems that it would suffice to use judiciously and in accordance with their instructions in order to neutralise the negative effects they could potentially have on individuals and society. In this regard, the advent in history of new techniques of expression and their effects on mental processes (memory, imagination, attention, reflection, etc.), creative thinking, intellectual productions, and cultural practices would merit more than ever being studied in depth, in order to be able to consider with a little more circumspection the scale and the nature of the changes that generative AI could produce in societies. Rather than yielding to the urgency and catastrophism that pit two camps against each other in a certain sense today — those who think one must rapidly change everything and react as quickly as possible for fear of being left behind, and those who think on the contrary that one must at all costs resist, precisely in the name of the preservation of a critical thinking more threatened than ever, as well as art, literature, philosophy, culture, and the ideal of emancipation inherited from the Enlightenment — one is compelled to note that we are faced with a dilemma that leaves both camps at an impasse.

On one side, we know that it will be difficult, not to say impossible, to stop the revolution in progress. And it is precisely in the name of change and the necessary adaptation it implies that the advocates of radical reforms call, often rightly, for the deconstruction and reform of existing institutions (schools, education, the world of work, social systems, political organisations, modes of democratic participation, etc.). On the other side, we are currently incapable of proposing documented and incontrovertible alternatives due to the brutality of the technological revolutions currently under way and their speed of propagation, which seems to take everyone by surprise. This dilemma is symptomatic of all the great industrial and technological revolutions, and takes on, in the current period, a particular acuity.

For my own part, I do not believe that forceful resistance to change and the prohibition of technologies whose proliferation seems inevitable constitute a viable solution in the long run. Nor do I naturally think that every technical advance should be welcomed uncritically, in a state of rapture that would cause us to forget the potential danger it contains. A more reasonable course is therefore to understand how these tools function, to think of them simultaneously as remedy and poison — the pharmakon — and to learn as far as possible to use them judiciously.