Search for...
Light Dark Auto
AI Posts
|
AI

ChatGPT-5 Prompting Best Practices

If your ChatGPT-5 results have gotten worse, it’s not because the model is less powerful - it's because the underlying architecture has changed, making old prompting techniques ineffective.

This article breaks down exactly why your prompts are failing and gives you five simple, actionable tips, from easy to advanced, to drastically improve your outputs.

Watch it in action

Resources

Why Your GPT-5 Prompts Are Failing

Before diving into the fixes, it's important to understand the two fundamental changes OpenAI made to GPT-5's architecture that render old prompting styles less effective.

1. Model Consolidation and the "Invisible Router"

Previously, users had access to many different models, but now there are only three main options. To manage this, OpenAI implemented an "invisible router" that automatically selects which model handles your request.

The problem is that the router doesn't always work well; if your prompt is simple, it may default to a faster, less powerful model to save on computing costs, leading to a lower-quality response.

2. Surgical Precision

OpenAI trained GPT-5 to be much better at following specific instructions, which is great for AI agents but can be a double-edged sword for the average user. While the model now adheres to prompts with extreme precision, it's significantly worse at guessing what you mean if your prompt is vague or poorly constructed. Sticking with old, less-structured prompting techniques will lead to worse results.

5 Tips to Fix Your ChatGPT-5 Prompts

Here are five tips, sorted from easiest to hardest, that you can use to adapt to these changes and improve your outputs.

1. Use Router Nudge Phrases (Low Effort)

You can force the invisible router to select a higher reasoning model by adding a simple phrase to the end of your prompt. This is especially useful for high-stakes tasks where you can't afford to miss second-order effects.

I’ve found three phrases that reliably trigger deeper reasoning:

  • "Think hard about this."
  • "Think deeply about this."
  • "Think carefully."

When deeper reasoning is triggered, you will see a "thinking" indicator, and the output will often include nuances and second-order effects that a standard response would miss.

2. Master The Verbosity Control (Low Effort)

Just as you can nudge the reasoning model, you can also use specific phrases to control the length of the output. After extensive testing, I found three phrases that work consistently for different needs.

  • For low-verbosity outputs (critical information only): Use the phrase, "Give me the bottom line in 100 words or less, use markdown for clarity and structure."
    • This is perfect for something like a concise Slack message to an executive.
  • For medium-verbosity outputs (key takeaways plus context): Use the phrase, "Aim for a concise 3-5 paragraph explanation."
    • This provides enough detail to explain context without losing the audience's attention.
  • For high-verbosity outputs (comprehensive documents): Use the phrase, "Provide a comprehensive and detailed breakdown (600–800 words)."
    • This is ideal for project briefs, research summaries, or other detailed materials.

3. Use OpenAI’s Prompt Optimizer (Medium Effort)

OpenAI has an official Prompt Optimizer tool that automatically rewrites your prompts to be more effective for GPT-5. It improves prompts by adding structure, eliminating vagueness, and including error handling.

While this tool requires a separate developer account with a payment method, there's a free workaround that works just as well.

You can use the following "meta-prompt" directly in ChatGPT-5 to have it optimize your initial prompt:

You are an expert prompt engineer specializing in creating prompts for AI language models, particularly `[model]`

Your task is to take my prompt and transform it into a well-crafted and effective prompt that will elicit optimal responses.

Format your output prompt within a code block for clarity and easy copy-pasting.

## Here’s my initial prompt:

4. Create an XML Sandwich (Medium Effort)

With GPT-5's need for precision, structuring your prompts is now critical. Using XML tags, which are like labeled containers, is a highly effective way to organize the different parts of your prompt.

Instead of providing a single wall of text, you explicitly label each component, such as <TASK>, <MY_RESUME>, and <JOB_DESCRIPTION>.

Here's an example for your next job interview:

<TASK>
Act as the hiring manager. Based on my resume and the job description, ask me the three most challenging behavioral interview questions I am likely to face.
</TASK>
<MY_RESUME>
[Paste Resume]
</MY_RESUME>
<JOB_DESCRIPTION>
[Paste Job Description]
</JOB_DESCRIPTION>

This structure helps the model better comprehend the task, which leads to dramatically better results.

5. Use The Perfection Loop (High Effort)

GPT-5 is excellent at critiquing and improving its own work, a feature OpenAI recommends we exploit.

The "Perfection Loop" involves instructing the model to create its own definition of a "world-class" response, grade its own work against that standard, and iterate internally until it produces a perfect result. This works best for complex, zero-to-one tasks like writing a finished document from scratch.

You don't need to write custom instructions every time. You can use this universal prompt at the end of your request:

"Before you respond, create an internal rubric for what defines a 'world-class' answer to my request. Then internally iterate on your work until it scores 10/10 against that rubric, and show me only the final, perfect output."

Finally, remember that these tips are not mutually exclusive, they can be stacked on top of each other for even better results!

If you enjoyed this...

You might also like: How to use Each Feature in ChatGPT!