How to prompt Gemini 3 Flash

Prompt Gemini 3 Flash with a short, direct brief: persona, output shape, and hard limits first, then only the facts the task needs. Google groups Flash with the other Gemini 3 models on structure. Critical constraints belong in the system instruction or at the start of the user prompt. Flash-specific notes on that page are about the current year, the knowledge cutoff, and staying inside provided context — not about writing a longer brief than Pro.

Flash still punishes padding. A dense bullet of facts beats a narrative warm-up. If the input is a long document, follow the long-context rule from the same guide: the document first, the question last. For an ordinary short task, do the opposite of a rambling brief. Say the goal, the format, and the limits in a few lines, and stop.

The prompt structure

Use short labeled blocks rather than XML:

Persona: Writing a sprint note for the product channel.

Output:
Bullets under "Shipped", "Slipped", "Next". Maximum 6 bullets. No preamble.

Task:
Write the update from the facts below only.

Context:
- Sprint 14 ended Friday.
- Shipped: CSV export, SSO login.
- Slipped: audit log (blocked on legal review).

For a quick task, a one-line persona and a one-line output rule are enough. Add context only when the facts are not already in the task. When the context is a long paste, move the task under it.

Density beats detail

• Prefer bullets over paragraphs in Context.
• Cut decorative background that doesn’t change the answer.
• Keep any persona to one line. A paragraph of backstory takes attention away from the task.
• Don’t add “show your reasoning” sections unless you actually want the reasoning in the output.

A short prompt gets you speed more reliably than asking the model to “be fast”.

Where should a Flash limit sit?

Put the limit in the opening block, with the persona and the output format. Google’s Gemini 3 section says those critical instructions belong at the beginning, including for Flash. A long pasted source is the case where the question moves to the end, after the material.

The Flash-only clauses on the same page are system-instruction lines, not extra essays. One tells the model the current year is 2026 when a query is time-sensitive. One states a knowledge cutoff of January 2025. One limits answers to facts in the user context. Add the clause that matches the job. Do not paste all three into a prompt that needs none of them.

Flash also wants the facts as bullets, not a warm-up paragraph. A persona longer than one line, a motivation for the task, and a restatement of the same limit in three blocks are the padding the short model does not convert into a better answer. Keep one copy of each limit in the opening block. If you are not asking for the reasoning itself, do not add a “show your work” section. The answer shape — columns, subject line, fenced code, language — still has to be named, because a dense prompt with no output line comes back as whatever format the model guesses.

Positive grounding

As with Pro, avoid a blanket “do not infer”. Tell Flash to calculate only from the figures provided and to mark missing numbers as [TBD]. Instead of “don’t make up links”, say “if no source is given, leave the link out”.

If you’re asking for live or current information Flash may not have, tell it to say when it can’t verify something and to answer from what it does know.

Name the output shape

• Tables: name the columns.
• Email: name Subject and Body.
• Code: name the language and whether you want a full file or a diff.
• Language: say which. If you write in another language and don’t specify, Flash may answer in English.

Why Flash prompts go wrong

• Pro-length briefs with nested explanations.
• Constraints placed first, so they get dropped.
• Blanket negatives that stop it deducing anything.
• Several tasks in one prompt without labels.

When does Gemini 3 Flash wander off a short brief?

The brief is Pro-length. Flash is still a Gemini 3 model: direct, structured, constraints stated up front. A multi-paragraph backstory before the task spends the opening on material that is not a critical instruction.

Time-sensitive questions with no year. The Flash notes say to tell the model it is 2026 when the query depends on the current day, and that the knowledge cutoff to declare is January 2025. Leaving both unstated asks the model to guess which calendar it is in.

Grounding left implied. The documented grounding clause limits the assistant to the user context and to facts provided there. “Be accurate” is not that clause. If the answer must not leave the draft, say so in the system instruction.

A long paste with the ask on line one. Same long-context rule as Pro: source first, question last. Density does not mean the question has to precede a transcript.

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

Same draft as the other guides. The rewrite only reorganises it. It does not add a product, an audience, or a word count that was not already there.

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

Persona: Announce a pricing change to people who already use the app.

Output: About 600 words. No hype. Name 1,000 one-off credits and the flat 13.99 a month.

Task: Write the post.

Context:
- Windows app.
- Readers already use it.
- Old price and new price are both in the task line.