LLM StateTree

DSL-driven StateTree AI generation for Unreal Engine 5. The agent writes .llmstate JSON, the editor turns it into a real UStateTree asset — tasks, conditions, Utility-AI considerations, parameter binding, all round-trippable.

Skill file: skills/llm-statetree/SKILL.md

---

When the Agent Loads It

The agent loads this skill when the user asks to:

• Create an AI behavior tree / state machine from a description (idle, patrol, chase, attack, etc.)
• Convert a behavior description into a .llmstate file
• Inspect or modify an existing .llmstate file
• Look up a StateTree node type (Task, Condition, Evaluator, Consideration)
• Bind parameters and references between nodes
• Use Utility-AI (scored action selection) instead of pure FSM

Triggers: .llmstate, StateTree, AI, behavior tree, FSM, utility AI, NPC, patrol, chase, attack, condition, task, evaluator, consideration.

---

Prerequisites

Requirement	Notes
UE 5.7+	Engine version
LLM StateTree plugin	Install from Fab or copy from this repo to Plugins/
Enabled in editor	Edit → Plugins → AI → LLM StateTree → restart editor
StateTree engine plugin	Built into UE 5.7; no extra install

The schema file (StateTreeNodeSchema.llmstateschema) is generated from UE reflection at editor startup — the agent reads the latest one to know what tasks/conditions/evaluators are available in your project.

---

Quick Example

Ask the agent:

"Make me a basic enemy AI: idle for 2 seconds, then chase the player, then attack when in range."

The agent produces a .llmstate file:

{
  "name": "SimpleAI",
  "version": "1.0",
  "parameters": [
    { "name": "IdleDuration", "type": "float", "default": 2.0 },
    { "name": "ChaseSpeed",   "type": "float", "default": 600.0 }
  ],
  "states": [
    {
      "name": "Idle",
      "type": "State",
      "tasks": [
        { "type": "StateTreeDelayTask", "name": "Wait",
          "properties": { "Duration": "${Param.IdleDuration}" } }
      ],
      "transitions": [
        { "trigger": "OnStateCompleted", "type": "GotoState", "target": "Chase" }
      ]
    },
    {
      "name": "Chase",
      "type": "State",
      "tasks": [
        { "type": "StateTreeMoveToTask", "name": "MoveToTarget",
          "properties": { "AcceptableRadius": 50.0 } }
      ],
      "transitions": [
        { "trigger": "OnStateCompleted", "type": "GotoState", "target": "Attack" }
      ]
    },
    {
      "name": "Attack",
      "type": "State",
      "tasks": [
        { "type": "StateTreeDelayTask", "name": "AttackCD",
          "properties": { "Duration": 0.5 } }
      ],
      "transitions": [
        { "trigger": "OnStateCompleted", "type": "GotoState", "target": "Idle" }
      ]
    }
  ]
}

Save it, open the LLM StateTree panel, click Generate — the StateTree asset is on disk.

---

What the Skill Knows

The full skill body contains:

• All node types — State, Group, Selector, Sequencer, plus all registered Tasks / Conditions / Evaluators
• Selection behaviors — TrySelectChildrenInOrder, TrySelectChildrenAtRandom, TrySelectChildrenByUtilityScore
• Transitions — trigger types (OnStateCompleted, OnCondition, OnEvent), Goto targets
• Parameters and bindings — ${Param.Name} reference syntax, context data binding
• Utility AI — Consideration node with scoring functions
• 5 example files — SimpleAI, GuardAI, UtilityAI, StealthAI, BossAI
• Auto-generated schema — points to the editor's "Generate Schema" button for the latest reflection data

---

Compiling the JSON into an Asset

# Manual: editor panel
Tools → LLM StateTree Panel → select .llmstate file → Generate

# CLI (with llm-easy-shell installed)
statetree generate /Game/AI/SimpleAI --from SimpleAI.llmstate

The StateTree asset opens in the StateTree editor.

---

Round-Trip

Already have a StateTree? Use Export to JSON in the editor panel to dump it to a .llmstate file. The agent edits it, you re-generate.

---

Related Skills

• llm-easy-shell → — drive the editor from the CLI: statetree generate, statetree inspect, etc.
• llm-easy-shell-lite → — read-only statetree inspect / list-types / export-schema.