The Prompt Moves That Actually Change What ChatGPT Gives You
Forget the 200-page prompt guides. Six moves changed my output more than any template, and most of them are about being specific about what you don't want.
I read the long prompt manuals so you don't have to, then ignored most of them. The handful of moves that changed my actual output were boring and specific, not clever.
Move one: give it a role and a reader. 'Explain like I'm a new hire' beats 'explain clearly' because it sets a bar the model can aim at. Vague instructions get vague answers and we both know it.
Move two: show, don't tell, with one example. A single input-output pair teaches the shape you want better than a paragraph of adjectives. I keep a tiny example file for the tasks I repeat.
Move three: ban the things you hate. Tell it 'no bullet lists, no intro sentence about being an AI,' and suddenly the output sounds like a person. The negative instruction does more work than the positive one.
Move four: ask for a draft then a critique. 'Now list what's wrong with your answer' catches the model hedging. The second pass is where the useful stuff hides.
ChatGPT (chatgpt.com)
Best for: anyone who writes, plans, or thinks with an LLM daily
Key features: conversation memory, custom instructions, file upload, image and voice input
Free plan: a free tier with daily limits on the stronger models
Pricing: Plus is around $20 a month for heavier use
Pros: versatile, fast, the default most people already have
Cons: it agrees with you unless told not to, which quietly rots your thinking
The thing nobody tells you: prompt engineering is mostly taste. The model does what you describe; your job is to describe the thing you actually wanted instead of the thing you said.
Frequently asked questions:
Q: Do I need a fancy framework like Chain-of-Thought?
A: Sometimes. For hard reasoning, 'think step by step' helps. For daily writing, plain specificity wins.
Q: Why does my prompt work once and fail later?
A: Models drift and context shifts. Keep examples and re-state constraints each session.
Q: Is prompt engineering still useful with smarter models?
A: Yes, but less. Better models forgive sloppy prompts; they don't read your mind.