A good prompt is a small interface, not a magic phrase.
You do not need secret words. You need to make the outcome, evidence, constraints, and finishing conditions easy for the model to see. Here is the method behind every template in the library.
Independent guide · General method
1. Outcome — name the job to be done
Start with the decision, artifact, or action you actually need — not the topic. “Compare three analytics tools and recommend one for a 5-person team” is a job. “Tell me about analytics” is a topic, and it produces a topic-shaped answer.
If you can state the outcome as “I will be able to ___ after reading this,” the model has a target to aim at.
2. Context — give only what changes the answer
Context is not everything you know; it is the subset that changes the response. Audience, constraints, and the current state of the work belong here. Irrelevant background dilutes the request and can pull the model off course.
- Who is this for? (sets depth and tone)
- What already exists? (repo, draft, dataset)
- What must not change? (protect existing work)
3. Constraints — set the boundaries
Constraints are where reliability comes from. Tell the model which sources are acceptable, how fresh the information must be, what tone to use, and — critically — what to do when it is unsure.
The single most useful constraint is an honesty clause: “If a claim cannot be sourced, say so instead of guessing.” It will not make the model perfect, but it reduces confident fabrication.
4. Finish line — define done and how it is checked
End every prompt by defining the output format and how quality is verified. A decision table, a diff with tests, a five-question quiz — a concrete finish line is easier to produce and easier for you to judge.
Match the model to the task
Heavier reasoning (multi-step research, multi-file code) tends to land better on the Pro tier; fast, well-scoped tasks (rewrites, quizzes, claim extraction) are cheaper and quick on Flash. Image and video are separate model families (Nano Banana, Veo). See the models overview for how we pick, and always check Google’s official model docs for current capabilities.
Verify before you trust
Every technique here reduces ambiguity; none of them make a model correct. Open the sources it names, run the tests it writes, and re-read rewrites for meaning drift. That is why our prompt pages carry a verification record and mark untested templates as Draft.
Put it together
Browse the prompt library to see the four parts applied to research, coding, writing, image, video, and learning tasks. Copy a template, replace the bracketed variables, and keep the verification instructions intact.
Not affiliated with or endorsed by Google. Gemini is a trademark of Google LLC. This guide describes a general prompting method and is not official Google documentation.