With Gemini 3, Google is introducing a new way of prompting, more direct but also more structured.
Who says new model also says new approach to the prompt. With Gemini 3, Google offers a better and simpler way to get even more precise answers. Launched on November 18 at the group level, the model was intensively tested by Google teams, who evaluated the quality of its responses according to different types of prompts. From these tests, Google has identified key principles, detailed in its (long, very long) documentation for developers. The JDN has retained the most relevant advice for anyone wishing to drastically improve the results obtained with Gemini 3.
1. Be direct, precise and unambiguous
This is the main change: like GPT-4.1 before, Gemini 3 is much more direct in its understanding of prompts. The model applies instructions with near-millimetre precision. This allows you to obtain results that are truly consistent with your request (prompt adhesion). The other side of the coin? The model is much more primary and requires precise, direct and unambiguous prompting. Google advises, at the same time, to be concise and to avoid instructions that are not very useful or performative.
Another new feature is that Gemini 3 must be prompted with beacons to maximize its effectiveness. A technique used for a long time at Anthropic which is starting to become widespread. Google insists, using XML or markdown (surely less efficient) allows the model to accurately understand the global context, priority instructions and variable context. Use one format and maintain it throughout the prompt. Finally, always in the spirit of avoiding errors, Google recommends absolutely banishing all ambiguities. Always explain and contextualize ambiguous actions, words or expressions by explaining them quickly to the model, without falling into the trap of digression, however.
2. Control the verbosity in the prompt
This is also something new, the verbosity of Gemini 3 must be managed directly in the prompt. By default, Gemini generates short probabilistic answers. To adapt the length of your response, explicitly specify the length level, format and structure of the expected response. Example: “Answer using between 500 and 600 words, without bullet points.”
3. Image, video and sound are as important as text
This is the third major lesson of this guide, and undoubtedly the most transformative: no longer ignore multimodality. Next-generation models, like Gemini, aren’t just text readers; they are natively multimodal. They were trained simultaneously on corpora of images, videos and audio, all processed within a single unified neural network.
The Google teams therefore insist on a crucial point: it is imperative to treat visual and audio modalities on a strict equal footing with the text. What does this mean in practice? Explicitly mention the visual or audio elements you are sending: “Analyze the graph at the top left”, “Compare the two diagrams”, “What does the person say in this video between 0:15 and 0:45?”. This direct reference allows the model to establish coherent links between the modalities.
For example, rather than saying “I’ve attached a screenshot of a dashboard. Can you analyze it?”, choose “Identify the three metrics that dropped the most this month on this dashboard”. The model implicitly understands that it must look for information in the visual content, just as it would in a table of textual data.
4. How to properly handle long contexts
Gemini 3 has a context of one million tokens. On paper, it is possible to send gigantic contexts to the model without any problem. The reality is more nuanced. Even if Gemini 3 has made great progress in managing long contexts (Gemini is one of the best LLM families on this point), the model can sometimes forget certain instructions (lost in the middle bias, in particular) and deviate from its initial objective.
To prevent the model from forgetting your instructions, when you use long contexts, Google recommends starting your prompt with the document, code base or other and ending it with the instructions. Another tip, when you give the model information or documents over several paragraphs, then use a transition sentence to recontextualize this mass of data. For example: “Based on the information above…” The goal is to make the model understand that it must use the data provided when executing the instructions.
5. Improve reasoning by planning
Plan and act! Last tip we take and it is taken directly from the last two years of research on LLMs. Planning and asking the model to think before responding greatly improves its accuracy and Gemini 3 is no exception. Google advises, for example, to explicitly ask the model to:
- break down the problem into subtasks
- check that all the information provided is complete
- create a specific plan to achieve the goal.
Example :
Avant de fournir la réponse finale, veuillez : 1. Décomposer l'objectif indiqué en sous-tâches distinctes. 2. Vérifier que les informations fournies sont complètes. 3. Etablir un plan structuré pour atteindre l'objectif
It is a good idea to ask the model to “think” about its answer before providing it to the user. Always with the aim of obtaining the best results.
Example :
Avant de renvoyer votre réponse finale, confrontez le contenu généré aux contraintes initiales de l'utilisateur. 1. Ai-je répondu à l'intention de l'utilisateur, et non simplement au sens littéral de ses mots ? 2. Le ton est-il fidèle au persona demandé ?
Final advice
The most important? Iterate gradually according to your specific use case. And above all, in production, rigorously monitor model drift: even the most efficient LLMs require continuous monitoring to maintain their effectiveness over time.
