Prompting

How to Write Prompts for Gemini (Google Style)

Gemini's edge is a very large context window and native multimodality. Feed it whole documents and images, then point it precisely at what you need.

Setuproll Team8 min read2026-06-20

Gemini's strengths shape how you should prompt it. It accepts very long inputs and reads images, audio and video alongside text. So instead of trimming your context down to fit, you can often hand it the whole thing and then ask a precise question against it.

Lead with the long context

When you paste a large document, a transcript or several files, put that material first and your instruction last. The model reads the context, then the ask. Label each chunk so you can refer back to it, the same way you would name sections of a brief.

prompt.txt

[DOCUMENT A: Q3 board deck]
...full text...

[DOCUMENT B: Q3 finance memo]
...full text...

Task: Using only DOCUMENT A and DOCUMENT B, list every claim
about revenue that appears in one document but not the other.
Cite which document each claim came from.

Anchor the answer to the source

With a big context, the risk is the model blending sources or drifting to its own knowledge. Tell it to use only the provided material and to cite which chunk each point came from. That keeps it grounded and makes the answer checkable.

Use multimodal inputs directly

You do not have to describe an image in words first. Attach it and ask about it. The same goes for a chart, a screenshot of an error, or a slide. Phrase the question as if the model can see the thing, because it can.

Input	A good multimodal ask
A screenshot of a UI	List every accessibility issue you can see in this screen.
A photo of a receipt	Extract the vendor, date and total as JSON.
A chart image	Describe the trend and name the two largest months.

Gemini - grounded extraction

You

[image: receipt.jpg] Extract vendor, date, total. Return JSON only.

Agent

{"vendor":"Blue Bottle","date":"2026-06-14","total":"18.40"}

Point at the image and pin the output shape.

Be explicit about scope

A large window is powerful but it is not a license to be vague. The more material you provide, the more important it is to scope the question tightly. Tell it which sections to consider, what to ignore, and how long the answer should be.

Name the inputs the answer should draw from.
Say what to exclude, such as appendices or boilerplate.
Ask for citations back to the labeled chunks.
Set a length budget so a huge context does not produce a huge answer.

Long context is not free attention

Even with a big window, a question buried in the middle of a giant paste can get diluted. Keep the instruction at the end, keep it short, and refer to your chunk labels so the model knows exactly where to look.

A Gemini prompt pattern

Paste or attach all context first, each piece labeled.
Add any image, audio or video inputs in place.
End with a single, scoped instruction.
Require grounding: use only the provided material, cite the source.
Set the output format and a length limit.

Google prompting guidanceGoogle's strategies for prompting Gemini, including long-context and multimodal tips.ai.google.dev

The mental shift with Gemini is from compressing your input to organizing it. Give it the whole picture, label the parts, then aim a precise question at exactly what you need.

0 Comments

Loading discussion...

← All guides