Prompt Engineering Tutorial and Prompt PracticeModels

Gemini 3

Gemini 3 Pro / Flash / Flash-Lite overview

CHAPTER PROMPT DECISION

01Prompt problem

Gemini 3 Pro / Flash / Flash-Lite overview

02Reviewable output

A reusable prompt, example set, evaluation record or safety rule with its task boundary preserved.

03Definition of done

Run at least one representative case, inspect the result and record what still requires human review.

TL;DR

Gemini 3 is Google's 2025 flagship: Pro offers 1,048,576-token input and 65,536-token output, strong multimodal + tool chain; Flash variants focus on high throughput and low cost.
Typical combo: Gemini 3 Pro for long document/table/screenshot analysis, Gemini 3 Flash for real-time interaction and batch processing, Flash-Lite for lowest latency scenarios.
Ecosystem: AI Studio free tier has Pro access (rate-limited), Vertex AI provides enterprise-grade security, data residency, and call monitoring.

Ultra-long context: product docs, contracts, logs, academic paper annotation and comparison.
Multimodal: UI screenshots, flowcharts, chart comprehension linked with code/config.
Data-intensive automation: tool calling + retrieval + structured output (JSON/tables).

Be explicit about output mode: set JSON schema for structured tasks, or write "review first, then output with validation" in system prompt.
Multimodal context: add text summary labels to images/tables, avoid uploading lots of irrelevant screenshots at once.
Long document chunking: summarize by section then consolidate. Have the model output "citation number + source paragraph" for traceability when needed.
Cost control: Flash/Flash-Lite for low latency/high concurrency; Pro for critical steps or proofreading.

vs ChatGPT 5.1: 5.1 is more mature in function calling and product experience; Gemini 3 offers better cost-efficiency for ultra-long context and image/table parsing.
vs Claude 4.5: Claude remains strong in safety and structured writing; Gemini 3 wins on large window size and visual/table detail.

Too many images dilute context: batch-generate text summaries first, then keep only key images.
JSON occasionally drops fields: lower temperature to 0-0.3, and require "self-check field completeness before output" in system prompt.
Tool calling chains too long: set max calling rounds and provide timeout fallback responses.