Knowledge Islands

← All guidance

Prompting Google Gemini 3

How to prompt Google's Gemini 3 line (3.1 Pro, 3 Flash, 3.5 Flash). Read the shared Prompting guides principles first; this guide carries only what is distinctive to Gemini 3. This guide is reference for cross-model work: building on, targeting, or evaluating Gemini 3.

When to reach for it

Gemini 3 is Google's frontier reasoning line, reached for long-context, multimodal, and agentic work when you are building on the Gemini stack or comparing against it. It is a reasoning model by default, which changes how you prompt it: the prompt-engineering scaffolding that older Gemini needed now works against it.

Prompting principles

  • Simplify — let it reason. Drop the chain-of-thought scaffolding you used to force reasoning on Gemini 2.5. Set thinking_level: "high" and give a simplified prompt without step-by-step instructions on how to think. Verbose, over-engineered prompts cause over-analysis.
  • Tune thinking_level. Values minimal / low / medium / high; default high (dynamic) on Pro and Flash. Lower it for latency; pair minimal/low with a "think silently" instruction for fast paths. Don't combine thinking_level with the legacy thinking_budget in one request — that 400s.
  • Keep temperature at 1.0. Google strongly recommends the default. Lowering it (a common habit for "deterministic" output) can cause looping or degraded performance on math/reasoning. Remove explicit low-temperature settings when migrating.
  • Steer verbosity up if needed. It defaults to terse, direct answers. For a conversational or chatty persona you must explicitly ask for it (e.g. "explain this as a friendly, talkative assistant").
  • Put instructions last for long context. For books, codebases, or long videos, place the question/instructions after the data, and anchor reasoning to it ("Based on the preceding information…").
  • One consistent format. Use XML or Markdown, not both — mixing the two confuses it. Keep structure uniform and define ambiguous terms explicitly.
  • Objective constraints, no manipulation. Give measurable constraints ("a summary of 3 sentences or less", not "a brief summary"). Strip emotional appeals, flattery, and artificial pressure — they don't help.
  • Avoid blanket negatives. Broad "do not infer" / "do not guess" instructions make it over-index and fail basic logic or arithmetic. For grounding, tell it explicitly to rely on the provided context and to flag missing data rather than fabricate — don't just forbid outside knowledge wholesale.
  • Don't overload one prompt. If a single pass asks it to summarise, extract, translate, and draft all at once, split into separate prompts.

API notes

  • Media resolution. Test media_resolution_high for dense document parsing. Gemini 3 defaults may raise token usage on PDFs but lower it on video; reduce media resolution explicitly if a request exceeds the context window.
  • Effort mapping. An OpenAI-style reasoning_effort auto-maps to thinking_level, easing cross-model harnesses.
  • Knowledge cutoff. January 2025 — for time-sensitive queries, state the current date in the system instruction and have it follow that when forming search queries.