Crimson Knight is a Souls style action RPG built in Unreal Engine 5, featuring deep melee combat, emergent enemy behavior, and a scalable AI pipeline.
I designed The Absolute to recognize repeated habits, then used readable telegraphs, counterplay windows, and playtests to make that adaptation understandable and beatable.
For The Absolute’s Phase 1 and 2 encounter, I paired a modular State Tree with prediction and repeated playtesting. The goal was to challenge habits without creating a boss that felt omniscient.
PRIMARY BOSS STATE TREE
Every move returns to Root: clear the previous state, read the current context, choose a response, execute it, and evaluate again.
Reset movement, rotation, and gravity so each decision begins from a reliable state.
Read the predicted next move from the n gram model. I tested n=2 and n=3, then chose n=3 because using the player’s previous two actions made the response feel more intentional while still leaving room to learn and react.
Gate the move pool by range, player health, stamina, and sword charges. Sword charges are the player’s primary way to heal, or they can be spent for additional damage, so reacting to them changes the risk of the next exchange.
Choose one scoped move, execute it, then return to Root. Short and long range pools keep the available responses relevant to the current spacing.
To introduce a new attack, we add its distance or context rule, place the ability in the selector, and let the shared flow handle cleanup, evaluation, execution, and the return to Root. That made the encounter highly modular and easy to extend without rebuilding the decision logic around every move.
Each change had a clear job: make the boss easier to read, more strategically meaningful, or fairer to learn without removing the pressure.
01 / READABILITY
The Absolute’s projectile was difficult to read and dodge.
With limited animation coverage, I slowed the opening, smoothly restored its speed, and added a windup sound that players could learn.
Players recognized the dodge window and found several successful evasion routes. The cue taught a response instead of simply lowering the challenge.
02 / COUNTERPLAY
Players could not tell when the n gram system was countering a repeated move, so they could not learn from the mistake.
I added eye and spear VFX whenever a counter began. A brief ignite effect reinforced the consequence if the counter connected.
Dodge spam could trigger a fiery stab toward the predicted dodge location. Repeated attacks could make her evade, deny damage, and reengage. Players could finally connect cause and effect.
03 / FAIRNESS
Early versions countered too many player behaviors, making Phase 1 feel relentless rather than intelligent.
I added a chance to enter an idle state where The Absolute circles and reassesses instead of countering immediately. That gives players room to read the fight and decide how to reengage.
Across two two-hour sessions with roughly 100 middle and high school students, the fight moved from about a 10 percent Phase 1 success rate to roughly 35 percent. Players described it as more balanced and achievable without losing the need to adapt.
The encounter worked because the decision logic, combat abilities, animation, and production plan were built to change together.
Built State Tree flows for patrol, engagement, stun recovery, and evasion, then connected the prediction model to counter selection.
Designed a modular Abilities and Effects pipeline for attributes, upgrades, status effects, and boss moves without hardwiring every variation.
Scoped sprints, assigned work, reviewed code, and translated combat and level design needs into systems the team could build and tune.