Most iteration is caused by a brief that under-specified something everyone assumed was obvious.
Include
- Reference images. More than one, from different angles. The single highest-value item.
- Dimensions, in real units, or a scale reference.
- The target. Engine, platform, polygon budget, texture budget.
- The role. Hero, standard or background — this sets the quality bar more than any other single fact.
- Style. Realistic, stylised, matching an existing set. Point at existing assets rather than describing.
- Deliverables. Formats, LODs, collision, rigged or not, which maps.
- Naming and folder conventions.
State the constraints that are not obvious
- Does it need to deform?
- Does it need to open, or have moving parts?
- Is it seen from all sides, or only the front?
- Is it instanced many times, or unique?
- Does it need to tile or align with other pieces?
Each of these changes the construction fundamentally, and each is routinely omitted because it is obvious to whoever wrote the brief.
The wear story
For anything with age or use, say where the wear is and why. "Weathered" produces uniform grime; "weathered, with the paint worn through at the handle and rust streaking below the hinges" produces something that reads as real.
Briefing a generator
The same discipline, differently expressed. A prompt is a brief with no one to ask clarifying questions, so ambiguity resolves to whatever the model considers typical.
Supplying reference images instead of description is the equivalent of showing rather than telling, and it is as effective with a generator as with a person.
Feedback
Be specific about what is wrong and why, rather than what to change. "The silhouette reads as a box from gameplay distance" is actionable; "make it more interesting" is not.