Prompt engineering
Prompt engineering is a dead end for art production.
Ever-longer prompts are a way of talking to one model, this month. They don’t hold a character on model, they don’t carry a composition, and they aren’t art direction anyone can review. Artists shouldn’t have to become prompt writers. In Mirva the artist draws and directs in plain words; Mirva writes the prompts. What decides the result is the context it works from and the skills it follows.
Why it’s a dead end
Words don’t carry a composition
“Tower on the left, arches behind” covers a thousand layouts. 26 prompts made 26 different caverns; one drawing made the one that was meant.
Every image starts from zero
A prompt doesn’t remember yesterday’s character. Consistency comes from context the model is given every time, not from piling on adjectives.
Prompts are model dialects
What works on one model misfires on the next, and models change monthly. Mirva works with many models and writes for each one, so you don’t have to.
A prompt isn’t authorship
A prompt alone generally doesn’t make you the author of the image. Your drawing and your decisions do. More on AI copyright →
What the artist wrote. What Mirva wrote.
The first render of a LENS character, from the board’s chat record. The artist pointed Mirva to a skill and typed six short lines, one of them with a typo. Mirva wrote the prompt.
The artist typed
Mirva wrote, for Nano Banana Pro
An alien character matching the input sketch design and pose exactly. Style: Polished concept art character render, confident economical brushwork, forms built from clean flat facets of color, strong clear value structure with decisive light and shadow, selective hard edges where surfaces turn and nice soft edges elsewhere, warm skin tones against a cooler restrained palette, sculptural chiseled rendering of face and fabric, crisp rim lighting, professional illustration finish. Simple 40% flat grey background, high detail.
- From the sketch (dotted)
- From the message, underlined (“color” read as the collar it could see)
- From the skill, written once (plain)
Later in the session, the artist typed
Mirva wrote three prompts for the views, each with the skill’s style and every color decided so far, then sent all four views to Meshy v7 Multi for the 3D model. The skill asks for “one saturated accent color”; by then the artist had settled on muted orange hems, so Mirva wrote “one desaturated accent color (desaturated orange)”.
Front (the input)
Side
Back
45°
What actually matters: context and skills
Context
What the image is of. The sketch fixes the composition and the pose. The board holds the references beside it. Lore writes down the characters and the world once: how they look, what is true about them. Mirva reads all of it before it generates.
Skills
How a kind of image is made: the style, the framing, the background, the steps, the checks. Written once as a document, with reference images if it needs them, and used by everyone on the board, the project or the whole team.
Kept current
Mirva reads a skill or lore page every time it applies it, so an edit takes effect on the next image. When a creative decision lands, it offers to record it there, so the next artist starts from it.
In a team, the pipeline has an owner
Game studios don’t ask every artist to write shaders. Technical artists build the tools once, and everyone uses them. Skills work the same way.
Artists
Sketch, paint over and direct in plain words. They pick a skill from a chip; they never see a prompt unless they want to.
Technical artists and pipeline engineers
Write the advanced skills: which model for which job, which references to attach, multi-step procedures (render, then the turnaround views, then the 3D model), and the house rules every result must follow. Scoped to a project or the whole team.
Art directors
Own the style and the lore. When they approve a decision, it goes into the document, and every artist’s next image follows it.
Questions
Do I need to learn prompt engineering to use Mirva?
No. You draw and direct in plain words, the way you would brief another artist. Mirva writes the prompts for each model, from your sketch, your message and the skills and lore you point it to.
Can I see the prompts Mirva writes?
Yes. Each generation step in Mirva’s chat opens to show the model it used and the prompt it sent, so nothing about a result is hidden.
Isn’t a skill just a saved prompt?
It starts as one, but Mirva reads it every time, adapts it to the image and the model, attaches its reference images, and follows its steps. Edit it, and the next image follows the edit.
Who should write skills?
Anyone who knows how a kind of image should be made. In a team, a technical artist or pipeline engineer usually writes the advanced ones, and the art director owns the style and the lore.