Gemini Prompts: 12 Copy-Ready Templates for Work and Images

The best Gemini prompts state the task, give the context, say what the output should look like, and name the audience. Everything below follows that pattern. Copy the templates, replace the [placeholders], and save the ones that work for you.

These templates cover everyday work tasks, photo editing and image generation. They're written for Google Gemini, but most will run in ChatGPT or Claude with small changes. Where Gemini behaves differently, we say so.

What makes a Gemini prompt work

Gemini does well when you tell it who the output is for and what shape it should take. It does worse when you hand it a vague request and hope. A prompt like "write an email about the delay" gets you a polite, generic email. A prompt that names the reader, the facts and the length gets you something you can send.

Five parts cover most cases:

• Role or situation: who is asking and why.

• Task: one clear verb. Summarise, rewrite, compare, draft, classify.

• Context: the facts, the document, the numbers. Paste them in.

• Format: bullets, a table in plain text, three options, under 150 words.

• Constraints: tone, things to avoid, what to do when information is missing.

You don't need all five every time. For a quick question, task plus format is enough. For anything you'll reuse, include all of them, because a reusable prompt gets run by you on a tired Friday afternoon, or by a colleague who didn't write it. The structure has to carry the intent when you're not there to explain it.

A before-and-after example

Here's a weak prompt for a task most people do weekly:

Summarise this meeting.

Gemini returns a paragraph of general recap. It mentions topics but buries who owns what. Now the version with structure:

You are helping me write meeting notes for a team that did not attend.

Summarise the transcript below into:

1. Decisions made (one line each)

2. Action items, formatted as: [owner] - [task] - [due date, or "no date set"]

3. Open questions that nobody resolved

Rules: do not add anything that is not in the transcript. If an owner or date is not stated, write "unassigned" or "no date set" instead of guessing. Keep the whole summary under 200 words.

Transcript:

[paste transcript]

The difference in the answer is practical. The first gives you prose you have to reread. The second gives you a list you can paste into Slack. The "do not guess" rule matters most: without it, models tend to invent owners to fill the gap.

Typical output from the structured version looks like this:

Decisions

- Launch moves from 12 May to 19 May.

- Pricing page copy stays as drafted.

Action items

- Priya - update the launch checklist - no date set

- unassigned - confirm the legal review slot - no date set

Open questions

- Who approves the final screenshots?

Notice the two "no date set" entries and the "unassigned" owner. That's the prompt doing its job. A model left alone would have written a confident Friday deadline that nobody agreed to, and you'd find out the following week.

Gemini prompts for everyday work

These are the ones people reuse most. Each is complete as written.

1. Cold outreach email

Write a cold email from me to [recipient role] at [company type].

About me: [your role and company, one sentence]

What I'm offering: [product or service, one sentence]

One specific reason I'm contacting them: [a recent post, launch, hiring signal]

Requirements: under 110 words, plain language, one question as the call to action, no flattery in the first line, no phrases like "I hope this finds you well". Give me two subject line options under 6 words each.

The specific reason is the part people skip, and it's the part that decides whether the email reads as written for one person. If you can't fill that line in, don't send the email yet.

2. Document summary for a busy reader

Read the document below and write a summary for [audience, e.g. a VP who has two minutes].

Start with the single most important takeaway in one sentence. Then give up to 5 bullets covering risks, costs and deadlines. End with the decision this reader needs to make, if any. If the document doesn't contain a decision, say so.

Document:

[paste text]

Change the audience and the bullets change with it. A finance lead and an engineer reading the same contract care about different lines, so run it twice with different audiences if both need a version.

3. Bug report from rough notes

Turn my rough notes into a bug report for engineers.

Use these headings: Summary, Steps to reproduce, Expected result, Actual result, Environment, Severity (low/medium/high with one-line reason).

If a detail is missing, list it under "Questions for the reporter" rather than inventing it.

Notes:

[paste notes]

4. Data cleanup and classification

Classify each row below into one of these categories: [category 1], [category 2], [category 3], Other.

Return the result as CSV with columns: row_number, category, confidence (high/medium/low). Do not explain your choices unless confidence is low; in that case add a short reason in a fourth column.

Rows:

[paste rows]

Spot-check a sample of rows before you trust the whole batch. Ten rows read by hand will tell you whether your categories are clear enough.

5. Rewrite for a different audience

Rewrite the text below for [new audience, e.g. customers who are not technical].

Keep every fact and number unchanged. Replace jargon with plain words. Keep it to roughly the same length. Then list any terms you simplified so I can check them.

Text:

[paste text]

6. Decision comparison

I'm choosing between [option A] and [option B] for [goal].

My constraints: [budget, timeline, team size].

What matters most, in order: [priority 1], [priority 2], [priority 3].

Compare them against those priorities only. Then give a recommendation in two sentences and name the one assumption that would change your answer.

That last line, naming the assumption that would flip the answer, is the most useful trick in the set. It exposes what the recommendation depends on.

Habits that help with Gemini specifically

Some advice depends on the model and the app you're using, and Gemini changes often, so treat this as what we've seen rather than a fixed rule.

Paste or attach the source instead of describing it. If you say "summarise our Q3 report" without the report, Gemini will produce something plausible-sounding and unconnected to your numbers. Attach the file or paste the text, and tell it to use only that material.

Put the instructions before the long material, then repeat the one rule that matters most after it. With a long transcript or document, a closing line such as "Remember: no guessing on owners or dates" keeps the constraint from getting lost.

Ask for a format you can check. Numbered lists, labelled headings and CSV are easy to scan for mistakes. A flowing paragraph hides them.

When the answer is close but wrong, reply with the correction instead of starting over. "Keep structure and tone, but cut section 2 to one sentence and add the budget figure" works better than pasting the whole prompt again with a vague tweak.

Gemini photo prompts: editing and analysing images

Gemini can take an uploaded photo as input, so photo prompts fall into two groups: asking about an image, and asking for an edited version. Which edits are available depends on the Gemini version and app you're using, so test with a throwaway photo first.

7. Describe and critique a photo

I've attached a photo I took of [subject]. Act as a photography editor.

Give me: (1) what works, in two bullets, (2) what weakens the image, in two bullets, focusing on composition, lighting and background, (3) three specific edits I could make in a standard photo editor, with rough slider directions such as "lower highlights" or "crop to a 4:5 ratio".

Be direct. Don't pad with praise.

8. Product photo edit

Using the attached product photo, keep the [product] exactly as it is: same shape, colour, logo and label. Replace the background with [a plain light grey studio backdrop / a wooden table by a window]. Keep the lighting soft and from the left, with a gentle shadow under the product.

The phrase "keep the product exactly as it is" does real work. Without it, edit requests often alter details of the subject along with the background. If a logo still drifts, name it explicitly in the unchanged list.

9. Extract text or data from a photo

The attached photo shows [a whiteboard / a receipt / a printed table]. Transcribe all legible text. Preserve the structure: use headings and bullets for a whiteboard, line items for a receipt, rows and columns for a table.

Mark anything you can't read with [unclear] instead of guessing.

The [unclear] marker is the same idea as the meeting notes rule. You'd rather see a gap than a confident wrong number on a receipt.

Gemini prompts for image generation

Image prompts work best when you describe the picture like a brief to a photographer or illustrator: subject, setting, style, light, framing. Short lists of adjectives give generic results. Specific details give distinct ones. "Beautiful, stunning, high quality" tells the model almost nothing; "early morning light from the left, shallow depth of field" tells it a lot.

10. Photoreal scene

A [subject, e.g. ceramic coffee cup with steam] on [surface, e.g. a worn oak table] beside a window in early morning. Shot as a close-up from a 45-degree angle, shallow depth of field, soft natural light from the left, muted warm colours, no text, no people. Aspect ratio [16:9 / 1:1 / 9:16].

11. Illustration for a blog or slide

A flat vector illustration of [concept, e.g. a team reviewing a roadmap], limited palette of [colour 1], [colour 2] and off-white, simple geometric shapes, generous empty space on the [left/right] for a headline, no text in the image, no gradients.

Leaving space for a headline is a small detail that saves a lot of cropping later. Asking for no text in the image is deliberate too: generated lettering is often garbled, and it's easier to add your own in a slide or design tool.

12. Iterating on a result

Keep the image as it is, with the same composition and subject. Change only: [one thing, e.g. the background to a night sky]. Do not change [the things you like, e.g. the character's clothing and pose].

Change one thing per round. If you ask for five changes at once, you'll usually lose something you liked.

Improving a weak prompt in three rounds

You don't need to get a prompt right on the first try. A quick loop does the job, and it works on any of the templates above.

1. Run it on a real input. Use an actual transcript, email or photo, not a toy example. Toy inputs hide the problems.

2. Mark what's wrong in the answer. Too long, wrong tone, invented detail, missing section. Name the specific flaw.

3. Fix it in the prompt, not just the chat. Add the rule that would have prevented the flaw, then run it again on the same input so you can compare.

Suppose your cold email comes back chatty and ends with two questions. The fix is a constraint, not a new prompt: add "one question only, and cut anything that doesn't serve it." Run it again on the same recipient details. If the result improves, keep the change. If it doesn't, undo it.

To get a second opinion on a draft, paste it in with a request like this:

Here is a prompt I use with Gemini: [paste prompt].

It produces [describe the problem, e.g. summaries that are too long and sometimes invent owners]. Rewrite the prompt to fix that. Keep my placeholders and structure, change as little as possible, and tell me what you changed and why.

Asking for the smallest possible change keeps you in control. Otherwise the model tends to rewrite everything and you can't tell which edit helped.

Making Gemini prompts reusable

A good prompt that lives in a chat history is a prompt you'll rewrite from scratch next month. Three habits fix that.

1. Turn the variable parts into placeholders. Anything that changes between uses, such as the audience, the pasted text or the length, becomes [placeholder]. The rest stays fixed.

2. Save it with a name that says what it does. "Meeting notes, decisions and owners" beats "Prompt 14".

3. Change it on purpose. When you tweak a prompt, keep the old version so you can tell whether the tweak helped.

This is where PromptPal helps. You can save these templates to your prompt library and keep older wording through prompt version control, so a bad edit is one click from undone. If a draft feels weak, the Chrome extension can refine it inline while you're in Gemini.

Where Gemini prompts go wrong

Most bad answers trace back to one of four causes.

• Missing context. Gemini can't see your company, your customer or your last meeting. Paste in what matters.

• No stated format. If you don't say "three bullets", you get whatever length it prefers.

• No permission to say "I don't know". Add a rule for missing information, as in the meeting notes example.

• Too many tasks in one prompt. Summarise, critique and rewrite in one go gives a muddy result. Split it into steps and feed each answer into the next.

Also check the output against your source. Models can sound confident about details that aren't in the material, especially with numbers and names. For anything going to a client, read it line by line.

If you work across several models, prompts that are structured this way tend to transfer better than clever phrasing. We cover that in how to get consistent results across Claude, ChatGPT and Gemini.

Next step

Browse the public prompt library for more role-based templates you can copy into Gemini, then save your own best versions so your team can reuse them.