How to write better Gemini prompts
Gemini will read far more than you are used to pasting, follows a named output structure closely, and answers from memory unless you tell it not to. Those three facts are most of what separates a good Gemini prompt from an average one.
Start from the same five ingredients
Nothing about Gemini replaces the fundamentals: a role, the context it cannot infer, one specific task, real constraints, and the format you want back. If those are missing, no model-specific trick rescues the prompt — the complete guide to prompt engineering covers them. Everything below is what Gemini rewards on top of that base.
Give it the whole thing, not a summary
Gemini's headline strength is how much it will read at once. Where you would normally paste an excerpt and hope you cut the right part, paste the entire document, transcript or thread and ask a narrow question about it. The model is better at finding the relevant paragraph than you are at guessing which one it will need.
Two habits make long inputs work. Put your instruction after the material rather than before it, so the last thing the model reads is the job. And label the material plainly — "Here is the transcript:" — so it never mistakes the reference for something you want rewritten.
summarize this meeting (pasting 3 bullet points of notes)
Below is the full transcript of a 50-minute product meeting. Read all of it, then list: (1) every decision that was actually made, (2) every open question, with who owns it, (3) anything two people disagreed about and did not resolve. Quote a line for each. Transcript: …
Ask for structure and you will get it
Gemini is unusually cooperative about output shape, and unusually likely to produce a wall of prose when you do not ask. Naming a table, a set of fields, or JSON with specific keys is the single cheapest upgrade to most Gemini prompts — and it makes the answer checkable, because a missing column is obvious where a missing sentence is not.
what should I consider when choosing a CRM
Compare HubSpot, Pipedrive and Zoho for a 6-person B2B sales team. Format as a markdown table with one row per tool and columns for: entry price per seat, the one thing it is best at, the main limitation at our size, and how hard it is to migrate off. Then one paragraph naming your pick and the strongest argument against it.
Say when the answer has to be current
Gemini can reach for live results, but it will happily answer from memory if your prompt does not signal that recency matters. Say so explicitly — "as of this month", "cite the source and its date", "if you cannot verify this, say so rather than estimating." That last clause matters more than people expect: it is the difference between a confident wrong number and a flagged gap.
Ask for the date beside every claim, not just the source. A correctly cited page from 2021 is still the wrong answer to a question about this year, and a citation without a date hides that.
Prompting with a screenshot or a photo
When you attach an image, the prompt's job changes: the model can already see it, so describing what is in the picture wastes the instruction. Spend the words on what you want done with it and what you are worried about instead.
what is wrong with this design
This is a screenshot of our checkout screen on a small phone. Assume the user is in a hurry and has used the app once before. List the three things most likely to make them abandon at this step, in order, and for each one name the specific element on the screen you mean.
Two mistakes specific to Gemini
- Trimming the input to be polite. Long context is the reason to reach for this model. Cutting a document down to what you think is relevant throws away the advantage and often removes the part that mattered.
- Accepting the first structure it picks. Left to itself, Gemini tends toward headed sections and generous prose. If you want six words in a cell, ask for six words in a cell.
Write the prompt without writing the prompt
Set the target to Gemini, type the rough version, and tap Enhance — Prompt AI Keyboard rebuilds it with the role, context, constraints and format Gemini responds to, in whatever app you are already typing in.
FAQ
How do I write a good Gemini prompt?
Start with the universal five — role, context, task, constraints, format — then use what Gemini is good at: paste the whole document rather than an excerpt, put your instruction after the material, and name the output shape explicitly. Gemini produces prose by default and a table only when asked.
Is prompting Gemini different from prompting ChatGPT?
The fundamentals transfer completely. The differences are that Gemini rewards very long inputs, follows an explicitly named output structure closely, and needs to be told when an answer has to be current rather than from memory.
How do I stop Gemini answering from memory when I need current information?
Say so in the prompt: ask for the answer as of today, require a source and its date beside every claim, and add an instruction to say so rather than estimate if something cannot be verified. Without that clause the model fills the gap confidently.