I build systems for a living. Give me an inventory grid, a save-state serializer, a pathfinding bug at 2 AM, and I'm genuinely happy. Ask me to describe what a character looks like and I produce things like 'she should feel dangerous but also kind of tired,' which is, technically, a sentence. My two-person art team was very patient about this. The first pass on our rogue protagonist, Vael, came back looking like a generic fantasy mannequin, because that is exactly the amount of information I had given them. They were working from fog.
The problem wasn't the artists, and it wasn't their talent. The problem was that I had a whole movie playing in my head and a working vocabulary of about eight adjectives to transmit it. We were one nod away from burning another two-week revision cycle on a guessing game, where I'd say 'no, more dangerous' and they'd dutifully add more spikes, and I'd say 'no, not like that,' and everyone would slowly lose the will to live. I needed to stop describing the movie and start showing a frame of it.
What I actually fed the prompt
I ran the Character Concept Sheet — Silhouette-First Game Hero or Villain Reference prompt on FLUX.2 [pro]. Character name Vael, role 'playable rogue protagonist.' Art style 'anime-influenced hand-drawn with thick clean outlines,' because that's our actual visual target rather than a stock fantasy look. World genre dark fantasy RPG. Color palette I typed in literally: slate-grey armor, deep teal accent cloth, worn brass hardware, pale skin. Aspect ratio 2:3. What came back was a real sheet — a front pose, a side profile, a close-up of the curved dagger, and a labeled swatch row across the bottom. It looked like the first page of an art bible from a studio that knew what it was doing.
It was not finished art, and it was never supposed to be. I want to be very clear about that, because the failure mode here is treating a generated sheet as the deliverable. It's a direction document. It's a thing my lead could point a finger at and say 'this, not that.' The whole value is that it converts my fog into something concrete enough to disagree with.
The conversation I finally got to have
I dropped the sheet into our team chat with zero commentary, just to see what would happen. Within a minute my lead artist wrote back: 'oh, I get it now — can we push the teal more cyan and segment the chest armor so it reads as plates instead of one slab?' That single message was worth the entire exercise. For the first time she wasn't interpreting a paragraph of my mood-words. She was reacting to an image, which is the thing artists are actually wired to do. We had a real, fast, specific back-and-forth in a single afternoon.
She kept the silhouette, killed the cape because it read as generic, kept the hooded asymmetry that gave Vael an off-center menace, and her second pass landed about 80 percent on my internal movie instead of the usual 30. The remaining 20 percent was now genuine artistic improvement on top of a shared foundation, not a desperate attempt to read my mind.
- First artist revision matched my intent about 80 percent, versus roughly 30 percent on the vibes-only brief
- Cut about two full revision cycles, which for our cadence is roughly three calendar weeks
- The sheet became our style bible — every new character now starts from one of these
- Vael's design locked in sprint 2 instead of dragging on into sprint 5 like the last one did
Why silhouette-first matters for games specifically
A game character has to survive being shrunk. In a dark dungeon, at a distance, or rendered as a 32-pixel sprite on a busy HUD, the only thing left of your beautiful design is the outline. That's why the prompt hammers on 'reads instantly at thumbnail size,' and FLUX.2 [pro] genuinely respects that constraint — the characters it produced had strong shape language instead of detail soup. My single dumbest, most reliable art-direction test is to squint at the thumbnail: if I can still tell who it is, the silhouette works. Vael passed the squint test. The original mannequin had not.
Honest caveat, because the failure modes are real: the first run gave Vael seven fingers on one hand, because of course it did, and a second run added a phantom buckle floating two inches off the belt. For finished art that would be a problem. For reference art it's a non-issue — I regenerated once, picked the cleanest of three, and moved on. I'm never shipping these exact pixels. I'm shipping the idea they communicate, and the idea was intact in all three.
If you're a dev who can't draw and you're tired of describing characters into a void, grab the Character Concept Sheet — Silhouette-First Game Hero or Villain Reference prompt on Prompt Dock and run it on your own half-formed character. Then pair it with my Game Asset Sheet — Unified Item Set for a 2D Game prompt to nail down the gear that hangs off them, and suddenly you have an actual visual language instead of eight tired adjectives and a patient art team.