All articles
Prompting · 5 min read

How to Write Prompts for ChatGPT That Get Better Outputs (Without Tricks)

The difference between a mediocre prompt and a great one is rarely cleverness. It is structure.

By Simple AI Prompt

The viral "secret prompts" lists overpromise. Real, repeatable improvements come from a handful of structural choices most users skip. Here are the ones that move outputs the most.

1. Give the model a job, not a topic

"Write about remote work" is a topic. "You are a workplace researcher writing a 600-word op-ed arguing remote work is misunderstood, for an audience of HR directors" is a job. Jobs win.

2. Specify the reader

Models match their register to the audience you name. "Explain APIs to a non-technical product manager" reads completely different from "explain APIs to a CS undergrad." Always name the reader.

3. Give it a frame to push against

Open-ended requests get safe, average answers. Add a constraint that forces a choice:

  • "Pick one and defend it. Do not list pros and cons."
  • "Argue the unpopular side."
  • "Limit to 3 ideas, ranked. Cut anything that does not earn its spot."

4. Treat format as an instruction

The model is great at following format instructions and bad at guessing them. "JSON with keys X, Y, Z." "Markdown table." "Five bullet points, each under 12 words." "No headings." Always specify.

5. Use two passes, not one

The single biggest free upgrade: ask for a draft, then ask for a critique and rewrite. The model genuinely improves its own work when asked to. Most users only ever ask for the draft.

6. Store what works

The 5th time you write essentially the same prompt, you have wasted 4 sessions. Save it as a template you can run with different inputs. That is what makes "prompt engineering" repeatable instead of artisanal.

That is exactly what the Loop Library gives you: a growing collection of templates that already encode the structural choices above, categorized by what you are trying to get done.

The shortest version of this article

Name the role. Name the reader. Specify the format. Add one hard constraint. Do a critique pass. Save what works.

Do that for a week and your "ChatGPT is mediocre" complaints disappear.