Write Gemini prompts with clear, labeled instructions

Why labeled sections suit Gemini

Google's prompt design guidance for Gemini leans on clear, direct instructions and on simple prefixes that mark the parts of a prompt, for example an input label and an output label. Uppercase labels followed by a colon do exactly that, without depending on how a given interface renders Markdown. The text stays readable whether you paste it into the Gemini app, a Google AI Studio system instruction field or an API call.

An Expert-level example (shortened) built from "explain compound interest to teenagers with a worked example":

ROLE:
You are a patient teacher who explains complex topics simply.

CONTEXT:
A 45-minute class for 15-year-olds with no finance background.

TASK:
Explain compound interest to teenagers with a worked example.

REQUIREMENTS:
- Start with a one-sentence summary
- Use an analogy or real-life example
- Finish with 3 quick self-check questions

EXAMPLES:
Example of the style I like: [paste a short example here, or delete this line]

QUALITY CHECK:
- Every requirement above is met
- No filler, clichés or generic statements
Before you start, ask me up to 3 short questions if any critical information is missing.

Refer to files, images and long inputs

Gemini handles more than text: you can attach documents, images, spreadsheets and, depending on where you use it, audio or video, and it can work with long inputs. The prompt is where you tell it what to do with them. Name each attachment in the CONTEXT section and say what role it plays: "The attached PDF is last year's report; the screenshot shows the new dashboard." Then make the TASK specific to those files, for example "compare section 3 of the report with the numbers in the screenshot".

When you attach several files, a short label for each one in the prompt helps Gemini keep them apart in its answer. If you need text from a photo or scan as editable input first, the image to text tool extracts it in your browser.

Examples and step-by-step tasks

Google's guidance also recommends few-shot prompts, meaning one or more examples of the input and output you expect, when you need a precise format. Use the EXAMPLES section in Expert mode for this: show one finished item, such as a product summary in your exact layout, and Gemini will tend to follow it. For complicated jobs, break the work into steps and send them one at a time, reviewing each answer before the next, rather than packing research, analysis and writing into one request.

Gemini prompting mistakes to avoid

• Vague time references: "recent" or "latest" means little. Give dates or ask Gemini to state the date of any information it uses.
• Unlabeled attachments: saying "look at this" when three files are attached leads to mixed-up answers.
• Contradicting the format: if you pick JSON output, do not also ask for an explanation in the same reply.
• Skipping the audience: Gemini adapts depth and vocabulary well when you name the reader; without it you get a middle-of-the-road answer.

Compare the same idea formatted for other assistants on the ChatGPT prompt generator, or use the universal version on the AI prompt generator.