This product or feature is subject to the "Pre-GA Offerings Terms" in the General Service Terms section of theService Specific Terms, and theAdditional Terms for Generative AI Preview Products. You can process personal data for this product or feature as outlined in theCloud Data Processing Addendum, subject to the obligations and restrictions described in the agreement under which you access Google Cloud. Pre-GA products and features are available "as is" and might have limited support. For more information, see thelaunch stage descriptions.

Prompting is a key part of working with any Gemini model and the newfeatures of Gemini 3 models can be prompted to help solve complexproblems and achieve other tasks, such as interpreting large amounts of text,solving complex mathematical problems, or even creating images and videos.

This guide provides a variety of prompting strategies to help you get the mostfrom Gemini 3 on Vertex AI for a variety of use cases.

Temperature tuning

For Gemini 3, we strongly recommend keeping the temperatureparameter at its default value of1.0.

Gemini 3's reasoning capabilities are optimized for the defaulttemperature setting and don't necessarily benefit from tuning temperature.Changing the temperature (setting it to less than1.0) may lead to unexpectedbehavior, looping, or degraded performance, particularly with complexmathematical or reasoning tasks.

Prompting strategies

The following sections describe a variety of prompting strategies that you canuse with Gemini 3 models.

Lowering response latency

For lower latency responses, try setting the thinking level toLOWand using system instructions likethink silently.

Distinguishing between deduction and external information

In some cases, providing open-ended system instructions likedo not infer ordo not guess may cause the model to over-index on that instruction andfail to perform basic logic or arithmetic or synthesize information foundin different parts of a document.

Rather than a large blanket negative constraint, tell the model explicitly touse the provided additional information or context for deductions and avoidusing outside knowledge.

Examples

What was the profit? Do not infer.

This instruction is ineffective because thedo not infer instruction is toobroad.

You are expected to perform calculations and logical deductions based strictlyon the provided text. Do not introduce external information.

Here, the instruction makes it clear that the model should use the providedcontext for calculations and reasoning.

Using split-step verification

When the model encounters a topic it doesn't have sufficient information for(such as an obscure place) or is asked to perform an action it doesn't havecapability for (such as accessing a specific live URL), it maygenerate seemingly plausible but incorrect information in an attempt to satisfythe request.

To avoid this, split the prompt into two steps: first, verify that theinformation or intended capability exists, then generate the answer based offof that information or capability.

Example

Verify with high confidence if you're able to access the New York Times home page.If you cannot verify, state 'No Info' and STOP. If verified, proceed to generatea response.Query: Summarize the headlines from The New York Times today.

Organizing important information and constraints

When dealing with sufficiently complex requests, the model may drop negativeconstraints (specific instructions on what not to do) or formatting orquantitative constraints (instructions like word counts) if they appear tooearly in the prompt.

To mitigate this, place your core request and most critical restrictions as thefinal line of your instruction. In particular, negative constraints should beplaced at the end of the instruction. A well-structured prompt might look likethis:

• [Context and source material]
• [Main task instructions]
• [Negative, formatting, and quantitative constraints]

Using personas

The model is designed to treat the persona it is assigned seriously and willsometimes ignore instructions in order to maintain adherence to the describedpersona. When using a persona with your prompts, review the persona that'sassigned to the model and avoid ambiguous situations.

Example

You are a data extractor. You are forbidden from clarifying, explaining, orexpanding terms. Output text exactly as it appears. Do not explain why.

Maintaining grounding

The model may use its own knowledge to answer your prompt, which might conflictwith any provided context. While the model is designed to be helpful, if youprovide a hypothetical scenario that contradicts real-world facts (promptingwith context such asCrabs are fictional and have never existed.), themodel may revert to its training data rather than your prompt to align yourrequest with its existing information.

If you need to work in context that isn't grounded in real-world information,explicitly state that the provided context is the only source of truth for thecurrent session.

Example

You are a strictly grounded assistant limited to the information provided in theUser Context. In your answers, rely **only** on the facts that are directlymentioned in that context. You must **not** access or utilize your own knowledgeor common sense to answer. Do not assume or infer from the provided facts;simply report them exactly as they appear. Your answer must be factual andfully truthful to the provided text, leaving absolutely no room for speculationor interpretation. Treat the provided context as the absolute limit of truth;any facts or details that are not directly mentioned in the context must beconsidered **completely untruthful** and **completely unsupported**. If theexact answer is not explicitly written in the context, you must state that theinformation is not available.

Synthesizing multiple sources of information

When information is presented in multiple places across a source of context, themodel can sometimes stop processing additional information after the firstrelevant match.

When working with large datasets, like entire books, codebases, or long videos,place your specific instructions or questions at the end of the prompt, afterthe data context. You can also anchor the model's reasoning to the provideddata by starting your question with a phrase likeBased on the entire document above....

Example instruction

Based on the entire document above, provide a comprehensive answer. Synthesizeall relevant information from the text that pertains to the question's scenario.

Steering output verbosity

By default, Gemini 3 models are less verbose and designed toprioritize providing direct and efficient answers.

If your use case requires a more conversational persona, you mustexplicitly steer the model to be chattier in the prompt.

Example instruction

Explain this as a friendly, talkative assistant.

What's next

• Learn more aboutGemini 3 Pro.
• Try theIntro to Gemini 3 Pro notebook tutorial.
• Learn aboutFunction calling.
• Learn aboutThinking.