A survey looks like the easy part of evaluation, which is exactly why it is so often done badly. Most surveys are written, without anyone meaning to, to confirm what the team already believes rather than to learn something that might surprise them. The result is data that feels reassuring and tells you almost nothing.
The usual culprits are familiar once you look for them. Leading questions that telegraph the desired answer. Double-barrelled items that ask two things at once, so you cannot tell which one the answer is about. Jargon that means something precise to you and nothing to the respondent. Scales that do not fit the question. And, above all, length; the forty-one-item survey nobody finishes tells you far less than the twelve-item one they do.
A survey is a conversation you cannot follow up on, so ask as if you only get one chance, because you do.
The principles that fix most of it are not complicated. One idea per question. Plain language a tired person understands on the first read. Balanced response options that make it as easy to disagree as to agree. Never ask for something you have no intention of acting on; it wastes the respondent's time and your credibility. And tell people, up front and honestly, that their answers are anonymous and what you will do with them.
Then pilot it with a few real people before it goes out, and watch where they hesitate or frown. The questions that confuse your test group will confuse everyone, and you only find them by asking a human being to try, before you ask three hundred.