← ClaudeAtlas

enemy-ai-frameworklisted

Architecture blueprint for game AI across genres: FSM/HSM, behavior trees, GOAP/HTN/utility AI, brain-intent-execution separation, perception, aggro, threat selection, attack tokens, lifecycle, AI LoD, group coordination, pathfinding, avoidance, crowds, and believability. Use when designing enemy AI, behavior trees, spawning, stealth/RTS/racing/sim AI, or when enemies swarm, flip targets, path badly, or feel robotic.
Firzus/agent-skills · ★ 1 · AI & Automation · score 72
Install: claude install-skill Firzus/agent-skills
# Enemy AI Framework Build the AI layer of a game. References: Genshin's GDC 2021 scalable AI (200+ designer-authored archetypes), the attack-token model (Doom 2016, God of War), F.E.A.R.'s GOAP and the "illusion of intelligence", Halo's archetype sandbox, souls-like aggro/leash. Excluded (covered elsewhere): boss phase data (`combat-system`), town-NPC schedules. ## The two architecture rules 1. **The brain decides; the body is the player's.** AI emits intents — `MoveTo(pos)`, `Attack(target, skillId)` — consumed by the **same character controller and combat system as the player**. Payoff: staggers/knockbacks work identically, no movement cheating, hitboxes implemented once, and you can plug an AI brain into the player character for soak tests. ("Have the AI press buttons.") 2. **Decision architecture by role, designer-authorability as the deciding criterion** (Genshin rejected one big BT for productivity): ``` FSM/HSM → lifecycle/context states (idle/patrol/combat/return, boss phases) Decision trees → action selection INSIDE a state (light, stateless, recomposable) or BT Utility/IAUS → continuous parametric choices (target, skill, position) GOAP / HTN → only when multi-step improvisation IS the fantasy ``` Genshin's shipped pipeline (per-frame, modular): `Sensing → Threat → Target Select → Reactions → Scripted → Group → Positioning` — modules recompose per archetype; designers author new enemies without engineers. ## Reference map | File