AIRules

NexusLink AI Workflow (IDE-side Rule)

Division of labor with MCP handshake instructions

• InitializeInstructions (UE handshake): capability routing table, meta-tool constraints, read/write naming.
• This Rule (IDE-side): enforce discover tools → find paths → read/write; forbid guessing cap names/paths; fill project-level pathFilter prefixes.

Source: Plugins/Developer/NexusLink/Resources/AIRules.mdc — diff generic sections into your project copy after plugin upgrades.

---

1. Four-step flow (required)

For UE assets / PIE / blueprints / Widget / material / behavior tree / GAS / Lua, call MCP this turn first, then answer:

Step	Tool	Notes
1 Discover capability	search_capabilities	Skip when cap name is known; otherwise query with 1–2 narrow words (e.g. blueprint graph, runtime actor, behavior tree); no business feature names, asset names, or over-broad single tokens (see §2)
2 Discover asset paths	search_asset	Requires assetType + functional pathFilter; top-level assets; typed search adds top-level recommendedGet/recommendedManage — prefer those in step 3
3 Read/write	get_asset_* / manage_asset_* / list_runtime_* etc.	assetPath only from assets[].path; prefer recommendedGet/recommendedManage; when lists include <k>_defaults, missing fields equal that value ({**defaults,**entry})
4 Persist	save_asset or manage saveToDisk=true	Required after disk asset edits; BP/ABP/WBP may also pass compile=true

Meta-tools search_capabilities, call_capability, submit_feedback must use MCP tools/call directly — not call_capability(capability="search_capabilities", …).

---

2. Prohibitions

• ❌ Invent a get_asset aggregate tool (correct: get_asset_<type>, e.g. get_asset_blueprint)
• ❌ Search search_capabilities with business terms (ChestPVE, Dice, Attack, etc.)
• ❌ Over-broad single-token search_capabilities queries: blueprint / asset / runtime / animation (causes search_overflow; add action or subdomain, e.g. blueprint variable, runtime actor)
• ❌ Guess /Game/... paths; no assetType=all + bare /Game/ without filters
• ❌ Call get_asset_* without search_asset first (missing assetPath)
• ❌ get_runtime_actor_* / set_runtime_actor_* without actorName (call list_runtime_actors first)
• ❌ exec_command without a non-empty command
• ❌ Runtime Lua tools (eval_runtime_lua, etc.) without PIE / UnLua ready
• ❌ Ignore _feedbackHint — call submit_feedback
• ❌ Expect success:true on write success (no error means success; only explicit success:false is failure)
• ❌ Pass assetPaths / actorNames / widgetNames to capabilities (cross-target via call_capability.calls[])
• ❌ manage top-level fields/rows/keys/widgets/ops or bare action synthesis (use operations[])
• ❌ Legacy keys: blueprintPath/newPath/ownerWidget/filePath/Lua path/classPath (see §3)