GPT-5 Prompt Best Practices

Good GPT-5 prompts provide enough information to complete the task while avoiding unnecessary text.

Start from the task

Start by clearly stating the objective. Instead of writing:

“Can you look at this and maybe improve it somehow?”

Make the prompt testable

write:

“Rewrite this customer-support response in a friendly professional tone. Keep it under 120 words and preserve all dates and prices.”

Measure before deployment

The second prompt gives the model a specific task, format, tone, and constraint.

For complex workflows, separate instructions into logical sections such as task, context, constraints, input, and required output. Structured prompts are easier to maintain and reduce the chance that important instructions are buried inside long paragraphs.

Practical checks before you proceed

Provide examples when the desired output format would otherwise be ambiguous. However, do not add ten examples where two demonstrate the pattern sufficiently.

Keep stable instructions separate from changing user information when your API architecture allows it. This makes prompts easier to manage and may improve opportunities for prompt caching.

Practical checks before you proceed

You should also avoid repeating instructions. Sending “be concise” five times does not necessarily make the result five times more concise—it mainly adds tokens.

For long-document tasks, provide relevant material rather than automatically including every document available to your application.

Practical checks before you proceed

Finally, evaluate prompts using real examples. A shorter prompt is only an improvement if output quality remains acceptable. Prompt optimization should target the best combination of reliability, latency, token usage, and cost rather than minimizing token count at all costs.