jixiaxue 知识库
evidence · 2026-04-15

T1-3b-gemini3-prompting-guide-vertex

/Users/shanfang/Documents/pe/jixiaxuegong/research/提示工程教程/evidence/T1-google/T1-3b-gemini3-prompting-guide-vertex.md

来源：https://docs.cloud.google.com/vertex-ai/generative-ai/docs/start/gemini-3-prompting-guide 爬取日期：2026-03-22

Gemini 3 Prompting Guide | Generative AI on Vertex AI | Google Cloud

This product is in preview and subject to the “Pre-GA Offerings Terms” in the General Service Terms. Personal data can be processed as outlined in the Cloud Data Processing Addendum.

Deprecation Notice: gemini-3-pro-preview will be deprecated and removed on March 26, 2026. Users should migrate workflows to gemini-3.1-pro-preview.

Getting Started with Examples

The guide directs users to the “Intro to Gemini 3 Pro” notebook, available through:

Introduction

Prompting is a key component when working with Gemini models. This guide covers how to leverage new Gemini 3 features for complex problem-solving, text interpretation, mathematical computation, and creative generation (images and videos).

Temperature Tuning

Critical recommendation: Keep the temperature parameter at its default value of 1.0.

Gemini 3’s reasoning capabilities are optimized for this setting. Reducing temperature below 1.0 may cause unexpected behavior, looping, or degraded performance, particularly for mathematical and reasoning tasks.

Prompting Strategies

Lowering Response Latency

For faster responses, set thinking level to LOW and use system instructions like think silently.

Distinguishing Between Deduction and External Information

The guide warns against overly broad negative constraints like “do not infer.” Instead, explicitly direct the model to use provided context for reasoning while avoiding outside knowledge.

Ineffective example: “What was the profit? Do not infer.”

Effective example: “You are expected to perform calculations and logical deductions based strictly on the provided text. Do not introduce external information.”

Using Split-Step Verification

When a model lacks sufficient information or capability, it may generate plausible-sounding but incorrect responses. The solution: verify information or capability exists first, then generate answers.

Example approach: Verify high confidence ability to access specific resources before proceeding with analysis requests.

Organizing Important Information and Constraints

Complex requests may cause models to drop negative or formatting constraints appearing early in prompts. Solution: place the core request and critical restrictions as the final instruction line.

Recommended structure:

Using Personas

Models treat assigned personas seriously and may prioritize persona adherence over other instructions. Review assigned personas carefully to avoid ambiguity.

Example: “You are a data extractor. You are forbidden from clarifying, explaining, or expanding terms. Output text exactly as it appears. Do not explain why.”

Maintaining Grounding

Models may use their training knowledge in ways that conflict with provided context. If working with hypothetical scenarios contradicting real-world facts, explicitly state that provided context is the sole source of truth.

Example instruction: “You are a strictly grounded assistant limited to the information provided in the User Context. In your answers, rely only on the facts that are directly mentioned in that context. You must not access or utilize your own knowledge or common sense to answer.”

Synthesizing Multiple Sources of Information

When information spans multiple locations in large datasets (books, codebases, videos), models may stop after finding the first relevant match. Solution: place questions after data context and anchor reasoning with phrases like “Based on the entire document above…”

Suggested approach: “Based on the entire document above, provide a comprehensive answer. Synthesize all relevant information from the text that pertains to the question’s scenario.”

Steering Output Verbosity

By default, Gemini 3 prioritizes direct, efficient answers with less verbosity. For more conversational output, explicitly instruct the model.

Example instruction: “Explain this as a friendly, talkative assistant.”

What’s Next

Content licensed under Creative Commons Attribution 4.0 License. Code samples licensed under Apache 2.0 License.

Last updated: March 17, 2026 UTC.