The gap between "the AI didn't get what I wanted" and "the AI built basically what I imagined" is almost always about how the request was structured, not about the tool's capability. A vague prompt gets a generic result because there was nothing specific to build against.
Genre alone isn't enough information
"A platformer" describes thousands of extremely different games. "A platformer where you can't jump, only dash horizontally, and lava rises one tile every ten seconds" describes one specific game a system can actually reason about and build toward. The second version isn't more words for the sake of it, every clause is a constraint that rules out wrong answers.
The four things worth naming every time
Core mechanic (what does the player do, moment to moment), win and lose conditions (what ends the game, in both directions), the setting's mood in a few concrete words rather than an adjective like "epic," and any hard constraint you already know you want (a specific control scheme, a specific visual reference). Leave any of these fully open and the generator will make a default choice for you, which is fine if you don't have a preference, and frustrating if you did but didn't say so.
Iterate by naming what's wrong, not just "try again"
"Make it better" gives a system nothing to correct toward. "The jump feels too floaty, and there should be a wall-slide" is something a generation loop can actually act on, because it names the specific gap between what exists and what you want. Treat each iteration like a bug report against a real build, not a vague vibe check, and you'll converge in fewer rounds.
What over-specifying costs you
There's a real ceiling here too: pile on enough contradictory or overly specific constraints and you can box a generator into a corner where nothing satisfies all of them, or you've effectively hand-designed the entire game yourself and generation isn't saving you any time. The sweet spot is specific about what matters to you and silent about what doesn't, which is exactly the discipline that also makes a human collaborator easier to work with.