ChatGPT prompts worth saving

The prompt that works is rarely the longest one or the cleverest one. It is the one written by someone who knows what they actually want.

That is also why most prompts you find online don't quite work for you. They were written by someone whose context was different from yours. The shape of a good prompt is contextual. What you can take from them is not the words — it is the patterns.

Below are the patterns I keep returning to, with prompts you can lift wholesale or adapt. You will not need all of them. You will probably build your own collection over time, and that is the real goal.

Set the role

A model with no role acts like the average of everything it has read. A model with a role acts like a person doing a job.

"Act as a copy editor reviewing this for clarity. Mark every sentence that could be tightened, and explain why."

"You are a senior product manager preparing for a difficult conversation with engineering. Help me think through what they will push back on, and what concessions are reasonable."

"Pretend you are a curious eleven-year-old. Ask me three questions about this topic that an expert would not think to ask."

Set the task with precision

Vague in, vague out. Specific in, specific out.

Bad: "Help me with my email." Better: "Rewrite this email to be warmer without losing the part where I am declining their request."

Bad: "Give me marketing ideas." Better: "Generate ten taglines under eight words for a halal-certified hotpot restaurant in Penang. Each should hint at one of: speed, abundance, family, freshness, value."

Bad: "Help me study for the exam." Better: "I have a marketing exam in three days covering Chapters 4-6 of Kotler. Quiz me with twenty questions of mixed format. After I answer all twenty, score me and tell me which chapters I am weakest on."

Set the form

"Give it to me as a table with columns for X, Y, and Z."

"Reply in three sentences. The first should disagree with me. The second should explain why. The third should propose what I should do instead."

"Use a numbered list. Each item should be a single sentence."

The model usually outputs in whatever shape it thinks is helpful. Often that shape is bullet points everywhere, regardless of whether bullets help. Tell it the shape you want.

Set the constraints

"Keep it under 200 words."

"Use plain English. No jargon. Assume the reader has a high-school education and is busy."

"Do not use the words 'leverage', 'synergy', or 'unlock'."

The constraints prompt is the one I underuse. It works particularly well for cutting AI's tendency toward soft, smooth, lifeless prose.

Show your work

The fastest way to get a useful response is to give the model your draft and ask it to react.

"Here is my draft. Tell me the three weakest paragraphs and why."

"I am thinking of replying like this: [draft]. What is the worst thing about this reply?"

"Here is my plan for the meeting tomorrow: [outline]. What am I missing?"

Asking AI to critique your thinking is more useful than asking it to think for you. You stay in the loop. It catches blind spots.

Iterate, don't restart

The first response is rarely the right one. That is normal. Edit instead of re-prompting from scratch.

"Make it shorter."

"Keep paragraph two. Replace paragraph one with something more direct. Drop paragraph three entirely."

"That is too formal. Try again with the tone of a smart friend explaining over coffee."

Most of the AI conversations I have are five or six exchanges, not one. Treating it as a one-shot oracle wastes its actual strength.

The real skill

The prompts above are not magic. They are examples of one underlying habit: writing your request the way you would write a brief to a sharp but unfamiliar new colleague. Someone who knows nothing about you, has never seen your work, will not ask for clarification, and will do exactly what you say.

Get good at that. The prompts come from there.