diff --git a/docs/plans/2026-03-08-npc-inner-monologue-design.md b/docs/plans/2026-03-08-npc-inner-monologue-design.md
new file mode 100644
index 0000000..06255df
--- /dev/null
+++ b/docs/plans/2026-03-08-npc-inner-monologue-design.md
@@ -0,0 +1,136 @@
+# Task 1.4: NPC Inner Monologue / Thought Bubbles
+
+## Overview
+
+Periodically generate inner thoughts for NPCs based on their needs, relationships, and recent events. Display as italic first-person text in the NPC info panel and mood-contextual emoji bubbles floating above sprites.
+
+## Architecture
+
+### Data Flow
+
+1. `ThoughtSystem` runs each tick, tracks per-NPC cooldowns and a global generation timer
+2. Every ~90s, collects followed NPCs + event-triggered NPCs into a batch
+3. Sends one batched LLM request for up to 8 NPCs
+4. Parses numbered responses, stores as `thought` component on each entity
+5. Broadcasts thoughts to clients via `npc-thought` socket event
+6. Client: info panel shows italic first-person text; sprite shows floating emoji
+
+### New ECS Component
+
+```typescript
+interface Thought {
+  text: string;   // First-person thought, e.g. "I could really use something to eat..."
+  emoji: string;  // Mood emoji, e.g. "🍖"
+  tick: number;   // When generated
+}
+```
+
+### New Server Systems/Services
+
+- **`ThoughtSystem`** — per-tick system managing timing, cooldowns, batch collection
+- **`ThoughtGenerator`** — batched LLM prompt builder + response parser (in `server/src/llm/`)
+
+### Triggers
+
+Each trigger respects a per-NPC cooldown of ~90s minimum.
+
+| Trigger | Condition |
+|---------|-----------|
+| Periodic | Every ~90s for followed NPCs (fallback if no event-driven thought recently) |
+| Need critical | Hunger or energy below threshold |
+| Post-interaction | Social outcome just fired |
+| Relationship tier change | Classification changed |
+| Idle | NPC wandering with nothing notable happening |
+
+### Batching Strategy
+
+- Collect up to 8 pending thought requests per cycle
+- Single LLM call with numbered NPC contexts (name, personality, state, recent events)
+- Parse numbered responses back to individual NPCs
+- If parsing fails for a line, skip that NPC (no crash, no retry)
+- Batch size, timer interval, and cooldown are configurable constants
+
+### Rate Budget
+
+- Target: ~20 thought requests/hour (leaving ~20/hour for backstories + narrations)
+- Batch size 8 means 20 requests can serve ~160 individual thoughts/hour
+- Per-NPC cooldown prevents any single NPC from dominating
+- Daily budget: ~480 thought requests/day well within 1000/day limit shared across all features
+- All timing constants configurable for tuning
+
+### Emoji Mapping
+
+Derived from thought context before LLM call:
+
+| Emoji | Condition |
+|-------|-----------|
+| 🍖 | Hungry (low hunger need) |
+| 😴 | Tired (low energy need) |
+| 😊 | Positive interaction / good mood |
+| 😤 | Negative interaction / frustrated |
+| 🤔 | Idle / pondering / curious personality |
+| 💭 | Generic fallback |
+
+### Socket Events
+
+**New server → client event:**
+
+- `npc-thought` — `{ entityId: EntityId, text: string, emoji: string }`
+
+Sent to all clients when a thought is generated. Clients decide whether to display based on local follow state and visibility.
+
+### Client Display
+
+**Info Panel (NpcInfoPanel.ts):**
+- New "Thought" section above Recent Events
+- Italic first-person text, e.g. *"I could really use something to eat..."*
+- Shows only the most recent thought, replaced when a new one arrives
+- Hidden when no thought exists for this NPC
+
+**Map Sprite (GameScene.ts):**
+- Small emoji rendered above NPC sprite head
+- Fade-in over ~0.5s, hold ~4s, fade-out over ~0.5s
+- Only one emoji at a time per NPC (new replaces old)
+
+### Prompt Template
+
+Batched template asking for multiple NPC thoughts in one request:
+
+```
+System: You write brief NPC inner thoughts in first person. One sentence each, grounded and specific. No purple prose. Respond with numbered lines matching the input.
+
+User:
+Generate a brief inner thought for each NPC:
+
+1. Bjorn — Personality: SOC:6, EMP:7, TMP:13, CUR:11. State: very hungry, just argued with Helga (rival).
+2. Helga — Personality: SOC:14, EMP:15, TMP:8, CUR:9. State: content, recently befriended Sven.
+3. Sven — Personality: SOC:10, EMP:12, TMP:10, CUR:16. State: idle, wandering.
+```
+
+Expected response:
+```
+1. My stomach is killing me... and that smug look on Helga's face isn't helping.
+2. It's nice having someone like Sven around — someone who actually listens.
+3. I wonder what's past those hills to the north...
+```
+
+### Priority Rules
+
+1. Followed NPCs always eligible for thoughts
+2. Event-triggered thoughts (need critical, tier change) eligible regardless of follow status — but only if the NPC is followed (to conserve budget)
+3. Non-followed NPCs do not generate thoughts (can revisit later)
+
+### Configuration
+
+All tuning values in a config object (similar to `relationshipConfig.ts`):
+
+```typescript
+export const thoughtConfig = {
+  periodicIntervalTicks: number;    // ~90s worth of ticks
+  perNpcCooldownTicks: number;      // ~90s worth of ticks
+  maxBatchSize: number;             // 8
+  needCriticalThreshold: number;    // e.g. 20 (out of 100)
+  emojiFadeDurationMs: number;      // 500
+  emojiHoldDurationMs: number;      // 4000
+};
+```