12 shortcuts that turn casual ChatGPT users into power users.
Most people use ChatGPT like a search engine. These 12 patterns turn it into something closer to a junior analyst who actually pays attention.
The gap between someone who uses ChatGPT well and someone who uses it badly isn't IQ. It's pattern memory. Power users have built a small library of moves they reach for automatically. Casual users type a question and hope.
Here are the 12 patterns that account for roughly all of the difference. Skim them, pick three, use them daily for a week. You'll never go back.
The 12 moves
| Pattern | What to type | What it gives you |
|---|---|---|
| Persona | "You're a [role]. Audience: [who]. Goal: [outcome]." | Output calibrated to the right register |
| Two-pass | First "draft a list of 10". Then "now pick the best 3 and explain why." | Quality through narrowing |
| Reverse it | "What's the strongest argument against this?" | Pushback you'd otherwise miss |
| Format hint | "Respond in a table with columns X, Y, Z." | Output you can act on |
| Show, don't tell | "Here's an example of what I want. Now do the same for [X]." | Style match by example |
| Constraint | "Under 100 words. No bullets. Use second person." | Forces tighter prose |
| Calibration | "How confident are you, 1-10? Where might you be wrong?" | Honest uncertainty |
| Sanity check | "What's a 7-year-old going to spot wrong with this?" | Catches obvious errors |
| Re-explain | "Now explain it like I'm 12." | Tests your own understanding |
| Steelman | "Make the strongest possible case for [position you disagree with]." | Sharper thinking on your own side |
| Edit, don't write | Write the draft yourself. Then "Edit this for clarity, keep my voice." | Better output than letting it write |
| Spec sheet | "Goal: X. Constraints: Y. Out of scope: Z. Now propose 3 approaches." | Useful work from vague intent |
The pattern most people miss
If you only adopt one habit, make it edit, don't write. The default behavior — "ChatGPT, write me a blog post about X" — produces ChatGPT-flavored prose. The bypass is to write the rough version yourself and ask the model to edit. The output sounds like you because the input sounded like you. The model preserves voice when it has voice to preserve.
The corollary: stop pasting in your draft and saying "improve this". Be specific about what you want changed. "Tighten the third paragraph", "remove the hedging in section 2", "find the one sentence I'd cut to make this 20% shorter". Specific edit requests produce specific edits. "Improve" produces generic Reddit-tone polishing.
The 30-second move that beats every fancy prompt
Tell the model what you've already tried and why it didn't work.
"I asked the model to do X. It produced Y. I want Z. The reason Y isn't Z is [reason]. Try again with this in mind."
This single move outperforms 80% of the elaborate prompt-engineering templates floating around. You're giving the model the failure mode to avoid. It's also faster than rewriting your original prompt from scratch.
What to stop doing
A few traps to drop.
- Stop saying "you're a world-class expert". Every prompt-engineering template has this. The model already tries to be helpful. Saying it's a world-class expert mostly inflates the prose style without improving the substance.
- Stop using "step by step" as magic words. It used to nudge models toward better reasoning. Most modern models do this by default for complex tasks. The phrase mostly adds clutter now.
- Stop one-shot prompting hard problems. Hard problems need two or three rounds, not one perfectly-crafted prompt. The conversation is the tool, not the message.
Build the small library of moves. Pick three to make automatic. The compound effect after a month is bigger than you'd guess.