What is instruction following?
Whether an AI response does exactly what the prompt asked, including format, length and constraints. Usually the first dimension rated in an evaluation task.
Instruction following measures how closely a response obeys the prompt. It covers the explicit requests ("three bullet points", "under 100 words", "in Spanish", "do not mention prices") and the implied ones, such as answering the question that was actually asked.
Why raters check it first
A response can be accurate and well written and still fail the user, because it ignored a constraint. Most rubrics rate instruction following as its own dimension, and many projects treat a broken explicit instruction as a major error regardless of how good the rest is.
How to rate it
Before reading the response, list every instruction in the prompt, including the small ones. Then check each one. Count words or items when the prompt sets a number; do not estimate. When a prompt's instructions conflict, check the guidelines for which one wins.
Example
The prompt asks for a product description in under 50 words with no exclamation marks. The response is 48 words and has one exclamation mark. In a side-by-side evaluation, it loses to a plainer response that follows both rules.