Clearly prompted constraints are the single biggest lever you can pull to improve the ideas reasoning models generate. Be clear about what constitutes success, and models will work until they've achieved it.