Write Shorter AI Prompts

Shorter AI prompts can reduce costs and make instructions easier for models to follow, but “shorter” should mean less unnecessary information, not less useful information.

Start from the task

Consider this verbose instruction:

“Please carefully review the text that I am going to provide below, and after reviewing it, I would like you to create for me a shortened version that contains the most important information and does not contain things that are unnecessary.”

Make the prompt testable

A shorter equivalent is:

“Summarize the text below in five bullets. Preserve key numbers and conclusions.”

Measure before deployment

The second version communicates more precise requirements using fewer words.

A useful prompt normally contains four elements:

Practical checks before you proceed

Task: What should the model do?

Context: What information does it need?

Practical checks before you proceed

Constraints: What rules must it follow?

Output: What should the result look like?

Practical checks before you proceed

You don't need headings for every small prompt, but thinking in these categories helps remove filler.

Avoid phrases that add politeness without adding meaning to an API instruction. You also usually don't need to repeatedly tell a modern model that it is an expert unless the role conveys useful task-specific context.

Practical checks before you proceed

Examples can be valuable, but use the smallest number required to communicate the expected pattern.

Finally, review system prompts regularly. Production prompts often accumulate obsolete instructions after months of changes.

Practical checks before you proceed

Shortening prompts improves token efficiency, but clarity remains more important than raw length. A clear 500-token prompt that consistently succeeds can be cheaper than a vague 100-token prompt that frequently requires retries.