Best AI Prompts for Gemini Workflows (2026)

Prompts for Gemini workflows#

Act as a Gemini API engineer. I will give you a diagram or screenshot of a user interface: [PASTE YOUR IMAGE DESCRIPTION OR ATTACH THE IMAGE]. First describe what you see in the image, then list the components and layout, then write React and Tailwind code for them. Use only [LIST YOUR ALLOWED LIBRARIES], and mark anything you had to guess as an assumption.

Write a system instruction for a Gemini code-refactoring assistant. It must set the role, state that the user's code and request come after the instructions, require changes to be listed file by file with a reason for each, forbid new dependencies, and tell the model to say so when a file it needs was not supplied. The target repository is [DESCRIBE YOUR REPOSITORY AND LANGUAGE].

Design a research-synthesis workflow that uses Gemini function calling. Write the function declarations (name, description, typed parameters, required list) for [LIST YOUR DATA SOURCES, e.g. a search endpoint and a PDF reader], explain which calls are independent and can run in parallel, and say which must happen in sequence. Add the system instruction that stops the model from answering without a tool result.

Write a Python example using the Gemini SDK that streams a response to a customer question and also handles a failed call. The question is [DESCRIBE YOUR USE CASE]. Show the system instruction, the streaming loop, and where your code must check the result before showing it to a customer. Do not invent method names: if you are unsure of an SDK call, say so.

Compare these product pages and return a feature matrix as JSON: [PASTE THE TEXT OF EACH PAGE, LABELLED A, B, C]. First write the JSON schema (enum values for categories, integer for counts, null allowed where a page is silent), then fill it in. If a page does not state a feature, return null instead of guessing.

What to add to any Gemini workflow prompt#

• The rules that must always hold (role, tone, output format) in the system instruction or at the start of the prompt, which is where Google's Gemini 3 guidance places critical instructions.
• A few varied examples with identical formatting. Google says prompts without few-shot examples are likely to be less effective, and warns that too many can cause overfitting.
• For long inputs, the documents first and your question last.
• Function declarations with a clear description, specific types (integer, string, enum) and a required list; Google suggests keeping the active set to 10 to 20 tools at most.
• A schema, not a sentence, when you need JSON: set the JSON MIME type and supply the schema, with description fields and enums.
• A line telling the model to say when the answer is not in the supplied context. Google's Gemini 3 Flash guidance includes such a clause.

A worked example#

System instruction: You are a contracts analyst. Answer only from the
documents supplied. If a term is not stated in them, return null for it.

[Contract A text, contract B text, contract C text]

Return JSON that matches the supplied schema. Fields: party_name (string),
renewal_months (integer or null), termination_notice_days (integer or null),
governing_law (enum: "England", "New York", "California", "other", or null).
Question: for each contract, what are the renewal and termination terms?

A good result for this task has:

• One object per contract, in the order supplied, with no fields outside the schema.
• Null, not a guess, for every term the contract does not state.
• Values that match the contract text when you check each one by hand, since valid JSON can still be wrong.
• Governing law only from the listed enum values.
• No claim about a clause the model could not have seen in the pasted text.

The full prompt library is on our prompt library page.

Limits of this guide#

This page covers the Gemini API as Google documents it, not the Gemini app or Google Workspace features. Google's docs say Gemini 4 and 3.6 or later ignore custom temperature, top_p and top_k values and that guidance differs between model versions, so check the model you use. We did not run these prompts, did not test SDK code, and could not confirm OpenAI's current prompting guide, because the archive copy was blocked.