How to prompt Gemini 3 Pro

Prompt Gemini 3 Pro by stating the task, the persona, and the output rules at the start, then putting long source material before the question that uses it. Google’s Gemini 3 notes ask for prompts that are direct and consistently structured, with critical constraints in the system instruction or at the beginning of the user prompt. For a long document, code file, or transcript, the material comes first and the question comes last, joined by a line such as “Based on the information above.”

That split is the part people miss. A short request is not the same shape as a long one. On a short request, the limits that must hold — length, persona, what the answer is allowed to use — belong up front, because Gemini 3 is described as following a clearly defined task and its constraints when those are stated directly. On a long request, burying the question in the middle of the source is the failure mode the same guide calls out: supply the context first, then ask. Default temperature stays at 1.0. Google says setting it below 1.0 can cause looping or weaker results on maths and reasoning. Gemini 3 is also direct unless you explicitly ask for a longer, more conversational answer.

The rest of this page is that ordering, applied to one rough prompt, plus the ways a Gemini brief usually falls over.

What makes Gemini 3 different

• Direct by default. Gemini 3 gives efficient, less verbose answers. For a chattier or more explanatory style, ask for it.
• Tuned for temperature 1.0. Google recommends leaving temperature at the default. Lowering it can cause looping or weaker reasoning on maths and logic.
• It takes personas seriously. An assigned persona can override other instructions, so choose it deliberately.

Where do the constraints go?

Google’s Gemini 3 guidance puts essential behavioral constraints, the persona, and the output format in the system instruction or at the very beginning of the user prompt. Use one delimiter style and stick to it. XML-style tags or Markdown headings both appear in that guidance. Do not mix them in a single prompt.

A long input is the exception to “instructions first.” Documents, code, and transcripts go first. The specific question goes at the end, after a transition such as “Based on the information above.”

Persona: You are writing an internal note for people who already use the product.

Output:
A blog post of about 600 words. No hype adjectives.

Task:
Announce that the Windows app is moving from 1,000 one-off credits to a flat €13.99 a month.

Context:
The readers already use the app. Name both prices. Lead with what changes for them.

If a limit really matters, keep it in that opening block rather than hoping a later reminder is the one that sticks. For a pasted report or a long thread, invert it: the paste first, the question last.

Ground it positively

A bare “do not infer” or “do not guess” can backfire: the model becomes so cautious it stops doing basic arithmetic or connecting facts across the document.

Instead, say what it should reason from: “Calculate and draw conclusions only from the provided figures. Don’t bring in outside information.” That keeps deduction on while keeping the answer grounded.

When you give it a hypothetical scenario that contradicts real-world facts, state that your context is the only source of truth for the session. Otherwise Gemini may fall back on what it learned in training.

Split-step verification

Asked about something obscure, or about something it can’t access such as a live web page, Gemini may produce a confident but wrong answer. Split the request in two:

1. First, confirm whether the information or capability is actually available.
2. If it isn’t, say so and stop. If it is, answer.

This small gate removes a large class of plausible-sounding fabrications.

Long inputs: question at the end

With large inputs such as books, codebases or long transcripts, put the material first and the question last. Anchor the question to the whole input, e.g. “Based on the entire document above…”. Gemini can otherwise stop at the first relevant passage and miss information later on.

Personas: keep them consistent with the task

Put the persona in the system instruction or at the start of the user prompt, next to the other critical rules. A persona that asks for one job and a task that asks for another is two instructions. Pick the role that matches the deliverable.

Faster responses

For latency-sensitive calls, use a low thinking level. A short, well-structured prompt does more for speed than asking the model to hurry.

When does a Gemini 3 Pro prompt ignore its own limits?

Four patterns show up against the current Gemini 3 notes.

Critical rules buried under a long preamble. Persona, output format, and behavioral constraints are supposed to lead. A page of background before the first rule makes the opening the least important part of the prompt.

A long document with the question in the middle. The guide’s long-context pattern is the reverse: all of the source first, then the question, then a bridge such as “Based on the information above.” A question typed above a pasted transcript asks the model to answer before it has been told to use the whole input.

Temperature turned down to “be more precise.” Google recommends leaving Gemini 3 sampling parameters at their defaults. Setting temperature below 1.0 is called out as a cause of looping or degraded performance on complex reasoning, not as a quality dial.

Two delimiter styles in one prompt. XML tags in the instructions and markdown headings in the pasted notes are two structures. Pick one and keep it.

What does the same rough prompt look like for Gemini 3 Pro?

This is a structure example, not a scored model output. The input is the same draft used across these guides:

write a blog post announcing our windows app is moving from 1000 one-off credits to a flat 13.99 a month plan, about 600 words, its for people already using it, dont be hypey

The opening block carries the rules. The facts from the draft stay in context. Nothing is added that the draft did not say.

Persona: Writing for people who already use the Windows app.

Output: About 600 words. No hype adjectives. Name both the old price and the new price.

Task: Write the announcement post.

Context:
- The app is for Windows.
- Pricing moves from 1,000 one-off credits to a flat 13.99 a month.
- The readers already use it.