Speaking better with AI is no longer a matter of style. It is the skill that separates average from excellent. A skill that connects the possible and the real.
In 2026, the world of generative artificial intelligence (Gen AI) experiences both consensus and paradox. The consensus is on the extraordinary capabilities of the models and their extensive fields of application. Capacities constantly reinforced by the improvement of models: new reasoning models, the advent of multimodal (integration of a wide variety of input and output data) and diversification of the “paths” of access to AI (online, open-source, local models, wrappers, aggregators, AI in SaaS, access with voice, embedded in hardware or even agentic).
But all this coexists with a major paradox: the capabilities of the models are under-exploited, in businesses and in individuals. Individuals only use a tiny part of the power of the models while companies struggle to achieve mass adoption. In short, there is an abysmal gap between the possible and the real. The numbers back up this diagnosis: according to the study “The GenAI Divide: State of AI in Business 2025” published by MIT (Massachusetts Institute of Technology), approximately 95% of generative AI pilot projects in business produced almost no measurable impact on the bottom line.
However, the authors assert, the culprit is not the quality of the models, but a “learning gap” — a learning deficit. A growing problem in communication between humans and AI. And where does this problem come from? Mainly difficulty mastering “machine language” or even “prompting”. A word, widely used, sometimes quite vague, but which covers a simple reality: how to improve communication between humans and AI? And why is it so important?
Because depending on how you talk to the AI (“inputs”) you will obtain radically different results (“outputs”). These results can be qualitative, creative and even prodigious! But also absurd (hallucinations), simplistic (AI slop) or even dangerous (misalignment)…I do not claim to be able to summarize in a single article all the possibilities, the issues and the new developments. But it is important to pave the way for new trends in prompting and discussions with LLMs…
Nine new maxims
- “Precision remains your best friend” A model does not read your thoughts: it completes the most likely request from what you give it. A vague prompt produces an average response, because the machine “averages” everything it has seen. The solution consists of 3 elements: a more precise context (who speaks, for whom, why), an exact task with clear constraints, and a well-detailed expected output format.
- “Don’t shine the lamp…In direct sunlight!” On a reasoning model (extended thinking, o3, Deep Think), the explicit “Chain-of-Thought” is useless or even harmful: the reasoning takes place in dedicated hidden tokens. Giving a clear brief, not a step-by-step procedure, is most effective. Breaking down a prompt to an extreme is a waste of time and increases the possibility of hallucination.
- “The most beautiful footprints quickly fade into the sand” Constraints drift over long sessions. Feeds critical rules back into each round via Attentive Reasoning Queries (ARQ): a checklist of questions that the model must process before responding. For example: you use a chatbot as a writing coach to rework a novel. At the beginning, you lay down your rules: “always keep my voice, don’t rewrite for me, propose but don’t decide, and stay on a familiar, not literary register.” After 20 messages, the model slowly drifts — it begins to rewrite entire paragraphs for you, in a more sustained style than yours. With ARQs, instead of just sending your request, you surround it with a mini-checklist: Before responding to me, first address these questions: 1. Do I suggest ideas, or do I rewrite for them? (I must suggest, not decide) 2. Does the register of my suggestions stick to his familiar voice, or am I sliding towards the literary? 3. Do I respect her vocabulary rather than imposing mine? Then give me your feedback. The model first briefly answers these three questions, then it gives you its feedback — which comes out naturally reframed, because it has just actively re-anchored itself to your rules.
- “Think the opposite way. To find the right one” Negative prompting frames the output as much as positive instructions: makes clear what you don’t want — no hedging, no parasitic formatting, no preamble. Example: Write a birthday message for my colleague. No emojis, no cutesy tone, no impersonal AI style.
- “The answers are bronze. The questions are silver. The grading is gold!” It has long been believed that the art lies in the response — formulating the right request, receiving the right text. It’s bronze: useful, but the most common metal. Then comes the money: learning to have the model write his questions himself, letting him refine the prompt before using it. It’s already better, but it’s still a crapshoot — you don’t know if the new formulation is really worth the old one. The gold is the rating. In 2026, meta prompting is no longer just about blindly rewriting a prompt: it measures it. We generate several candidates, we evaluate their output, we keep the best, we start again. Without this evaluation loop, optimizing a prompt is just an opinion; with her, it’s a science. The centerpiece is not the well-asked question, it is the scale which decides between two rival questions.
- “We don’t just talk with words” For a long time, talking to an AI meant typing text. Those days are over: in 2026, she sees your photos, reads your PDFs, deciphers a screenshot, listens to a voice memo. The word is no longer the only language — the image, the sound and the document have entered the conversation. But handing him a file with a vague “what is that?” is no longer enough, just as you don’t hand an X-ray to a doctor without saying where it hurts. The right reflex is to frame: what you want her to look at, why, and in what form you expect her response. The more precise your request, the more precise his look is too.
- “Compare. Compare. Compare” No model is best everywhere. One excels at coding, the other at writing, a third at long reasoning or analyzing documents… This truth is all the more obvious as the different models multiply, mix, cross… Get in the habit of submitting the same prompt to several “frontier models” and comparing: you will learn as much about the models as you do about your own request. This competition reveals the blind spots, biases and strengths of each person — and prevents you from locking yourself into a single supplier. In 2026, knowing which model to use for which task, and why, has become a skill in its own right.
- “A spark in the wrong place can burn the whole forest.” In an agentic system, a bad prompt does not produce a bad response but a bad action that propagates throughout the pipeline. The system prompt becomes the most critical engineering artifact. The parade consists of two reflexes. First, tend to the first ember: the instructions that govern early decisions (classify, sort, choose a tool) are the most critical, because they are the ones whose errors have the most time to propagate. Then, install firewalls: checkpoints where the agent must verify or validate a decision before building on it — so that an error remains an isolated misstep instead of becoming a chain reaction.
- “No longer think in shots. Think in loops. And again in loops” The old reflex of prompting is archery: you adjust your request, you aim for a long time, you shoot, and you hope to hit the target on the first shot. We refine the perfect prompt, press Enter, and wait for the response. This mentality makes sense when the model only produces text — but it has become a liability in 2026. Why? Because the tasks we entrust to AI today are rarely solved in a single throw. Writing a solid report, debugging code, planning a project, analyzing a document: these are back-and-forths by nature. The good prompt no longer looks for the single arrow, it prepares the archer to shoot again — it anticipates a cycle that repeats itself. This cycle has a recognizable form, in four stages which go in circles: think (the model reasons about what to do), act (it produces something – a draft, a piece of code, a decision), observe (we look at the result, or the model itself evaluates it), start again (we correct the situation and we leave for another round). This is exactly the pattern called ReAct when an agent executes it with tools, and Reflection when the model self-evaluates its output and redoes it as long as it is below a certain quality threshold. But the idea goes well beyond agents: it applies to your everyday conversation.
These maxims are only part of the tremendous discoveries that humans are making, and will make, regarding this new art that now connects us to AI. I have good news! Everything you are going to do to better prompt is work… perfectly human! You will read books, find ideas on forums or media, talk with AI experts, have fascinating debates with colleagues, try slightly crazy experiments (and yes!), try to combine tools, carefully analyze the quality of your answers, compare the differences in models or even refine your results/creations by sharpening your critical thinking. In short, AI will make you smarter, more creative, more human. And this is undoubtedly, for the moment, one of the best ways of working with you 🙂
PS: Essential reading – How to talk to AI (And how not to) – Jamie Bartlett
