Prompt Engineering Cheat Sheet

A strong prompt names a role, a context, a task, a constraint, and an output. Skip any part and the model will quietly guess. You do not need special jargon to do this. Write who it is helping, what it must produce, and what it must not invent. If you can point to those five pieces, the prompt is ready to run.

Put the context the model cannot know before the task, not after a long preamble about your life. A paragraph of background is useful. A dump of every email is not. Say what is already decided so the draft does not reopen it. Say what is still open so the draft does not pretend you chose. Two short lists beat one foggy story.

Name the output in concrete terms: a table with three columns, a 90-word email, or five interview questions. Asking the model to make it good is not an output. If you need a tone, name a reader and a relationship, such as a polite note to a manager you know. Ask it to show assumptions at the end. Those assumptions are the first thing you will check.

Constraints are how you stop fluent extras. Limit length, ban invented numbers, and require a label when the model is unsure. Negative constraints matter as much as positive ones. Telling it not to add case studies can save a proposal. Telling it not to write your conclusion can save a student. If a constraint is missing, expect a generic voice.

When the answer is too long, do not merely say shorter and hope. Give a ceiling, such as 120 words or six bullets. Ask it to keep the decision and cut the throat-clearing. When the answer is too vague, ask for one example using only facts you supplied. When the tone is wrong, paste one sentence you wrote and say to match that, not a celebrity.

If a fact looks invented, quote the sentence and ask which words came from your notes. Tell it to replace anything else with a clear gap marker. That move is better than asking whether it is sure, which often produces a more confident repeat. Then you still verify outside the chat. The cheat sheet is a way to ask. It is not a substitute for looking.

Use an example when format matters more than cleverness. One short sample of the shape you want is enough. Two samples help when you need a contrast, such as a formal line and a plain line. Do not paste a whole brand book. Say which parts of the example are required and which were only flavor. Otherwise the model copies jokes you did not mean to keep.

Debug in this order: outcome, missing context, missing constraint, then tone. Rewrite only the broken piece and run it again in the same chat if the context is still valid. Start a new chat when earlier wrong instructions are being obeyed. Write down the sentence that fixed it. That sentence belongs in your library, not only in yesterday's thread.

The checklist pattern asks for coverage, not prose. Give the criteria and tell the model to mark each one met, missing, or unclear, using only the text you paste. This is how you review a draft, a lesson plan, or a proposal without inviting new ideas. Follow with a human pass on every item marked met that would be costly if wrong.

The options pattern forces a choice into the open. Ask for three options, the tradeoff of each, and what information would change the recommendation. Forbid a single favorite until you say which tradeoff you value. This keeps Gemini from hiding a preference inside smooth prose. You still decide. Redact confidential numbers before you run the pattern.

The critic pattern is a second prompt, not a mood. Paste your draft and ask what a skeptical reader would misunderstand, what claim lacks support, and what can be cut. Tell it not to rewrite the whole piece unless you ask. Apply the fixes yourself so the voice stays yours. If the critic adds new facts, delete them before they sneak into the text.

The example pattern is easy to overdo. One clean sample teaches format. A pile of samples teaches the model to mimic flaws, including outdated claims. Mark the example as a format sample, not as a source of truth. Then give the real notes underneath. End by asking it to list any detail that did not come from the notes.

Read the prompt once as if you were busy and literal. Would you know what to produce? Would you know what not to invent? If either answer is no, fix the prompt before you blame the model. Add the output shape or the missing ban. This short read prevents a clever chat that never stated the job.

Match the effort to the stakes. A lunch joke can be scanned. A price, a medical claim, a grade, or a quote needs a source you open yourself. Put that rule in the prompt when the stakes are high: list claims, leave blanks, and do not cite anything you did not provide. The cheat sheet will not make Gemini careful on its own. Your instruction and your check do that.

Keep a personal top ten prompts in a document you control, not only inside a chat product. For each one, store the job, the prompt, and a note about what you must verify. Revisit them when they start producing clutter. This course is independent of Google, so your library should survive a UI change. Check the current screen only for where to paste, not for what to say.

When you are stuck, switch from generating to diagnosing. Paste the bad answer and ask what in the prompt allowed it. You will often find a missing ban or a missing reader. Fix that line and rerun. Do not collect prompts like souvenirs. A short sheet you actually use beats a long sheet you admired once.