How to Write Better Prompts: 5 Changes That Make an Immediate Difference

Most people rewrite their entire prompt when ChatGPT gives them something off-target. The fix is usually much smaller than that.

After two years of working with language models and writing a book about prompt engineering, I've noticed that the same handful of missing elements account for most of the bad outputs people get. These are not exotic techniques. They are structural habits that make prompts more specific without making them longer.

Here are five changes that consistently improve outputs without starting from scratch.

1. Name the audience

Adding "explain this to a non-technical manager" or "write this for a first-year student" does more than setting a reading level. It activates a completely different set of vocabulary, analogies, and assumed knowledge in the model's output.

Before: "Explain how neural networks work."

After: "Explain how neural networks work to a marketing professional with no technical background. Use one analogy from everyday life."

The second prompt produces something immediately useful for that specific person. The first produces a response that is technically correct but calibrated for nobody in particular.

This works because the model has been trained on enormous amounts of text written for different audiences. Specifying the audience shifts which part of that training it draws on. You are not adding a constraint; you are pointing at a different area of the model's knowledge.

2. Specify the format before the task

Most people describe what they want and leave the format implicit. The model then picks whatever format feels plausible, which is often not what you needed.

Put the format instruction first:

"In three bullet points, each one sentence: explain the main differences between supervised and unsupervised learning."

Stating format first means the model structures its response before generating content, rather than shaping content and then trying to fit it into a format. The result is tighter and more consistent.

This applies to any output format: numbered lists, tables, JSON, code with inline comments, a paragraph under 100 words. The more specific the format, the less the model has to decide on its own.

3. Add a word limit

Without a limit, the model defaults to thoroughness. With a limit, it has to prioritize. A 100-word limit on a summary produces a tighter, more useful output than no limit, because the constraint forces the model to identify what actually matters.

"Summarize the following article in 100 words. Focus on the actionable takeaways."

This is particularly useful for summaries, descriptions, and any output that tends to run long. A 200-word limit is not the same as asking for "a brief summary." The number makes the constraint real. The model takes it seriously in a way that adjectives like "brief" or "concise" do not.

If you find yourself editing the model's output to cut it down, add a word limit to the prompt instead. You will get a shorter first draft that still contains the most important content.

4. Tell the model what not to do

Negative constraints are underused. They are often the fastest way to eliminate a recurring problem in your outputs.

If the model keeps adding disclaimers: "Do not include any disclaimers or caveats."

If it keeps hedging: "Write with confidence. Do not use phrases like 'it's worth noting' or 'it's important to consider'."

If it keeps repeating points: "Each point should be new information. Do not repeat what was said in a previous bullet."

If it keeps opening with a definition when you don't need one: "Do not begin with a definition of the term. Assume the reader already knows what it is."

Negative constraints work because they eliminate specific output patterns without constraining the content itself. The model still has freedom in what it says; it just can't say it in the ways you've ruled out.

5. Iterate on the response, not the prompt

When the first response is 70% of what you need, the fastest path to a good output is editing the existing response, not rewriting the prompt. Follow up directly:

"Good. Now make the second paragraph half as long and cut any jargon. Keep the first and third paragraphs as they are."

"The explanation is correct but too abstract. Add one concrete example after the second point."

"Change the tone to be more direct. Remove the phrases 'it's important to' and 'you should consider'."

This approach is faster, and it gives the model specific feedback about what is and is not working. A rewritten prompt removes all the context you have built up in the conversation. A targeted follow-up preserves what was good and fixes what wasn't.

Conversations are stateful. Use that. The model can adjust its previous response with far more precision than it can generate a new one from scratch.

Why these five

These changes address the most common reasons a prompt produces an unhelpful response: the audience is undefined, the format is unspecified, the length is unconstrained, a recurring problem has not been excluded, or the user starts over when a small correction would have worked.

None of them require learning new syntax. They are habits that become automatic after a few weeks of applying them consistently.

If you want to understand the principles behind these habits, and build a more systematic approach to writing prompts across different tasks, Practical Prompt Engineering covers the full range of techniques from zero-shot and few-shot prompting to chain-of-thought and prompt templates.