a reaction arcade i founded, designed and art-directed

the game

project lw is a mobile reaction arcade. the hero stands in the centre of an arena with four sectors around him and an enemy waiting in each one. an attack is telegraphed, you tap that sector inside the block window, and the round only ends when you run out of health.

i founded the project, wrote the design document and ran the team — and the art direction stayed with me. what follows is the loop as it stands in the alpha build.

the whole game is one question asked faster and faster: which sector, and can you still get there.

the tempo curve does the rest. spawn intervals shorten, the block window narrows, inverted and combo enemies arrive, and a boss lands every 60–90 seconds — guaranteed, so randomness never assembles an impossible sequence.

heroes, charges, difficulty

heroes are the meta layer. you start with a base hero and buy the rest with soft currency; each one changes how a round is survived rather than how fast it is won.

abilities run on charges. before the round you choose how many to carry — three per ability, six if the hero has two — and they burn whether or not you spend them. charges come back slowly on a timer, from daily challenges, or from a rewarded video.

the interesting decision happens in the preparation menu, not in the round.

difficulty scales on four dials: spawn interval, block window, enemy mix and boss cadence. if a player keeps dying inside the first thirty seconds the window widens a little; if they survive comfortably, it tightens.

finding the style

the prototype was grey boxes and a working block rule. nobody on the team drew, so the look was mine to decide — and a style is only a set of decisions somebody has to make.

the style was found on one character. the finished dark hunter is on the left, in motion; on the right are the four directions we tried before settling — a cold monochrome pass, chibi proportions, and two siblings drawn to the same rules: the veiled lancer and the watcher.

three directions for one hero

one character settles a look; a roster needs range. before anything went to a 3d artist i ran three deliberately different directions, giving up randomness as each pass moved closer to production. concept work earns its keep by removing uncertainty before modeling starts — it is not there to produce final illustrations.

one — silhouette first

so each pass had to leave a modeler less to guess at: silhouette, proportions, materials, costume logic and visual direction, decided before anyone commits time to geometry.[1]

i used generative ai as a rapid visual exploration tool, moving through three deliberately different art directions and increasing the level of control with each pass.

the first direction was intentionally rough. the character was a veiled lancer: narrow proportions, a hooded face, a long double-bladed weapon and an angular silhouette. at this stage i deliberately avoided solving every costume detail. the important questions were much larger:

does the character read immediately? does the weapon feel like part of the identity? does the silhouette communicate danger before we see the face?

the prompt therefore concentrated on large-form decisions:

the result was useful precisely because it was not over-resolved. the branch established an archetype and shape language quickly enough to decide whether the direction deserved further development.

this mirrors a common character-production principle: recognizable silhouettes and large forms should survive before detail is allowed to take over.[2]

two — how far the style can move

the second exploration went in almost the opposite direction. instead of a restrained physical threat, i explored an eccentric arcane character built around:

the interesting problem here was visual hierarchy. glowing effects are extremely easy to overuse. once everything glows, nothing actually feels magical anymore.

so the effects were treated as accents rather than decoration. the face and major silhouette still had to read first; light trails, symbols and smaller effects were there to support the character's story.

this phase also introduced multiple style references rather than relying entirely on text. that made the process less about finding increasingly imaginative adjectives and more about building a reference-guided visual constraint system.

three — something a 3d artist can read

the final branch had a different objective: less illustration-first, more production-aware.

i explored a female knight with a more grounded fantasy treatment and a much more explicit design specification. instead of describing only a mood, the generation was broken into separate layers:

this was also the stage where i combined style references and character references. the reference image was no longer simply inspiration. it became part of the control mechanism.

exploration vs control

one of the useful lessons from the project was that the right amount of randomness changes during the process.

early exploration benefits from variation. if the system is locked too tightly, it only produces increasingly polished versions of the first idea. later refinement has the opposite requirement: once a promising character appears, excessive variation starts destroying the decisions that were already made.

i therefore treated generation settings as part of the design process:

more freedom while searching. more constraints while refining.

for the later character branch, controlled variation was combined with style and character references so that the model could still offer alternatives without repeatedly reinventing the entire design.

Detail Masonry Slider

arenas and the handoff

every arena in the hub started as a pile of my concept sheets — drawings, colour call-outs, notes on what must survive production. in the centre is what the artist built from them; around it, the six concepts it grew out of.

from a pretty image to a model brief

this project also clarified the difference between concept art and key art.

a beautiful three-quarter illustration can communicate mood extremely well while still leaving a 3d artist guessing about the back of a costume, material separation or how individual pieces connect. traditional game-production workflows solve this with turnarounds, orthographic views and additional construction references.[1][3]

the exploratory images in this project were therefore used primarily to answer the earlier questions:

which archetype? which silhouette? which materials? which visual language? which character details are worth preserving?

once those decisions are stable, the next production step is not “generate an even prettier image.” it is to convert the selected design into cleaner references for modeling.

that is where the concept artist on the team took over. he received the whole exploration set, the selected direction and the final character style sheet, and drew production concept art from it — front, back and side views, separated pieces, weapon and hand studies, variants of the same base.

by the time these sheets were drawn the style was no longer in question, which is why they could spend their effort on construction instead of mood.

what did not work

what i learned

the project began as a fast way to create character ideas. the more useful lesson was that ai exploration needs different rules at different stages of production.

the strongest workflow became:

explore broadly → identify the visual idea → lock the important decisions → refine only what still needs solving.

generative ai accelerated the search. the actual design work was deciding what deserved to survive the search.

an art director’s job is not to draw the game. it is to make the decisions that let others draw it.

the build is in alpha. i will leave updates here soon.

references

1. 1proce, v. & sineni, c. next-generation of character concept art: creating photo-real people who don't exist. gdc, 2006.
2. 2bachynski, l. bringing orphea to heroes of the storm. gdc, 2019.
3. 3game character production from a 2d concept. 80 level — character-production breakdown discussing silhouette and early 3d blockout.