We found (in the world of building Hyperscan, a high-performance regular expression matching library) that asking users what features they wanted was a disaster

We found (in the world of building Hyperscan, a high-performance regular expression matching library) that asking users what features they wanted was a disaster. You have to ask them what they want to achieve. Their suggestions as to how to do it were almost always terrible: partly, they don't have the domain knowledge to know what's going to be simple and practical vs complex and nasty. Also, they usually have a comparatively clear picture of what they want to achieve, but their notions of how best to break that into "your problem (library)" vs "our problem (app)" are often very naive - you wind up owning way too much of the problem (usually), or occasionally going in the other direction and not really getting enough context information into your library.