How to Prompt Gemini: The Complete Guide

Gemini processes images, audio, video, and PDFs as first-class input — not via text descriptions or OCR. Embed media directly in your prompt for the model to analyse natively. This is fundamentally different from GPT-4o vision (images only) or Claude (images + PDFs).

// API request with mixed media
{
  contents: [
    { text: "Compare these two product designs. For each:\n1. Visual appeal (1-10)\n2. Usability issues\n3. Accessibility concerns" },
    { inlineData: { mimeType: "image/png", data: "<base64_design_A>" } },
    { inlineData: { mimeType: "image/png", data: "<base64_design_B>" } }
  ]
}

// Also supports:
// - Video: "Watch this 5-min demo and summarise key features"
// - Audio: "Transcribe and analyse sentiment of this call"
// - PDF: "Extract all financial data from this 80-page report"

💡 Don't describe images to Gemini — embed them directly. "Analyse this image" + embedded image produces far better results than a text description of the image.

Gemini 2.5 Pro supports 1M tokens (2M in preview) — 5-10× larger than GPT-4o (128K) or Claude (200K). This enables entirely new prompt patterns: whole codebases, full document repositories, hours of audio, and comprehensive data analysis in a single prompt.

// Whole-codebase analysis
"Here is our entire React application (847 files, ~400K tokens).\n\nPerform a comprehensive code review:\n1. Architecture issues\n2. Security vulnerabilities\n3. Performance bottlenecks\n4. Dependency risks\n\nPrioritise by severity and provide fix recommendations."

// Multi-document analysis
"Here are 12 quarterly reports (2023-2025).\nIdentify:\n- Revenue trends across all quarters\n- Recurring themes in risk factors\n- Strategic pivots between years"

💡 Even with 1M tokens, place the most important instructions at the beginning and end of the prompt — attention is strongest at the boundaries (primacy and recency effects).

Enable the Google Search tool to let Gemini verify responses against live web data. The model decides when to search — factual queries, recent events, and data-dependent answers trigger grounding automatically. Responses include grounding citations for verification.

// Enable search grounding
{
  model: "gemini-2.5-pro",
  tools: [{ googleSearch: {} }],
  contents: [{
    text: "What are the latest OWASP LLM Top 10 vulnerabilities for 2026? Include specific examples and remediation for each."
  }]
}

// Response includes:
// - Grounded answer with citations
// - Search queries used
// - Source URLs for verification

💡 Use search grounding for any query where accuracy depends on current information — pricing, regulations, recent research, competitor analysis, news.

Gemini applies safety filters that may block outputs containing violence, sexual content, hate speech, or dangerous activities — even in legitimate contexts (security research, medical content, creative writing). Configure safety settings explicitly for your use case.

// Configure safety settings
{
  safetySettings: [
    { category: "HARM_CATEGORY_DANGEROUS_CONTENT",
      threshold: "BLOCK_ONLY_HIGH" },
    { category: "HARM_CATEGORY_HARASSMENT",
      threshold: "BLOCK_MEDIUM_AND_ABOVE" },
    { category: "HARM_CATEGORY_SEXUALLY_EXPLICIT",
      threshold: "BLOCK_MEDIUM_AND_ABOVE" },
    { category: "HARM_CATEGORY_HATE_SPEECH",
      threshold: "BLOCK_MEDIUM_AND_ABOVE" }
  ]
}

// Thresholds: BLOCK_NONE, BLOCK_ONLY_HIGH,
//             BLOCK_MEDIUM_AND_ABOVE, BLOCK_LOW_AND_ABOVE

💡 If responses are being unexpectedly blocked, check which safety category triggered. For security research or medical content, set BLOCK_ONLY_HIGH for the relevant category.

Unlike GPT (strict system→user hierarchy) and Claude (XML-parsed structure), Gemini treats instructions more uniformly. System instructions exist but don't override user content as strongly. Place critical rules alongside your task, not just in the system prompt.

// System instruction
{
  systemInstruction: "You are a data analyst. Always respond with structured tables. Use GBP for all currencies."
}

// User prompt — reinforce key rules here too
{
  text: "Analyse the following sales data.\n\nIMPORTANT: All currency values must be in GBP. Format as a markdown table with columns: Product, Revenue, Growth%, Status.\n\n[data...]"
}

💡 Repeat the most important constraints in both system instruction AND user prompt. Gemini responds well to emphasis — "IMPORTANT:" and "CRITICAL:" prefixes strengthen instruction following.

Few-shot extraction minimizes context window usage vs zero-shot verbose.

3 well-crafted few-shot examples (150 tokens) outperform a 600-token verbose instruction block, saving 75% on input costs per request.

Without concise few-shot examples, developers write lengthy prose instructions that consume 4x more tokens for equivalent or inferior output quality.

JSON Schema enforcement eliminates parse errors.

OpenAI structured outputs with JSON Schema achieve 99.9% schema adherence vs <70% with unconstrained generation — a 30x reduction in parse failures.

Without schema enforcement, every 1M requests generate 300K+ malformed responses requiring retries, error handling, and downstream data corruption.

Chain-of-thought prompting improves complex reasoning accuracy.

Adding 'Let's think step by step' improves accuracy on GSM8K math benchmarks from 17.7% to 78.7% — a 4.4x improvement on multi-step reasoning tasks.

Without chain-of-thought, models attempt to produce answers in a single leap, failing on problems requiring intermediate steps.