diff --git a/skills/varg-video-generation/references/prompting.md b/skills/varg-video-generation/references/prompting.md
new file mode 100644
index 0000000..369c21e
--- /dev/null
+++ b/skills/varg-video-generation/references/prompting.md
@@ -0,0 +1,123 @@
+# Prompting Guide for varg Video Generation
+
+Use this when video outputs look generic, inconsistent, or lack motion. Rich prompts dramatically improve Wan/Kling results.
+
+## The 4D Formula
+
+Every video prompt should cover four dimensions:
+
+1. Subject: who/what is the focus
+2. Scene: where it happens and what it looks like
+3. Motion: how things move and with what intensity
+4. Cinematic controls: how the camera and lighting behave
+
+Template:
+
+```text
+[Subject] + [Scene] + [Motion] + [Cinematic controls]
+```
+
+## Subject
+
+Describe appearance, materials, clothing, and expression.
+
+Weak: "a woman"
+
+Stronger:
+
+```text
+a young woman with shoulder-length auburn hair, wearing a worn leather jacket,
+green eyes reflecting determination, jaw set with quiet resolve
+```
+
+## Scene
+
+Layer in environment, lighting, and mood.
+
+Weak: "in a city"
+
+Stronger:
+
+```text
+on a rain-slicked Tokyo street at night, neon signs reflecting pink and blue on
+wet pavement, steam rising from a nearby food stall, warm light spilling from a
+ramen shop window
+```
+
+## Motion
+
+Be explicit about movement, amplitude, and pacing.
+
+Weak: "walking"
+
+Stronger:
+
+```text
+walking slowly through the crowd, shoulders hunched against the rain, one hand
+raised to shield her eyes, droplets catching the neon light as they fall
+```
+
+## Cinematic Controls
+
+Use film language to control results:
+
+- Camera movement: push-in, pull-out, pan, tilt, tracking shot, static camera
+- Shot type: extreme close-up, close-up, medium, wide, over-the-shoulder
+- Composition: rule of thirds, centered, symmetrical, low angle, high angle
+- Lighting: soft light, hard light, rim light, backlight, golden hour, moonlight
+
+Example cinematic suffix:
+
+```text
+medium close-up, golden hour, dramatic rim light from behind, slow push-in,
+shallow depth of field, cinematic, film grain, anamorphic lens flare
+```
+
+## Character Consistency
+
+Generate a character once and reuse it across clips.
+
+```tsx
+import { render, Render, Clip, Image, Video } from "vargai/react";
+import { fal } from "vargai/ai";
+
+const hero = Image({
+  prompt:
+    "portrait, centered composition. warrior princess with crimson hair, emerald eyes, battle-worn silver armor, a thin scar across her left cheek",
+  model: fal.imageModel("flux-schnell"),
+  aspectRatio: "9:16",
+});
+
+await render(
+  <Render width={1080} height={1920}>
+    <Clip duration={4}>
+      <Video
+        model={fal.videoModel("wan-2.5")}
+        prompt={{
+          images: [hero],
+          text:
+            "wide shot on a windswept cliff, cape billowing in the wind, dust particles in golden light, slow camera pull-out",
+        }}
+      />
+    </Clip>
+    <Clip duration={3}>
+      <Video
+        model={fal.videoModel("wan-2.5")}
+        prompt={{
+          images: [hero],
+          text:
+            "medium close-up, she draws her sword in one fluid motion, blade catches the dying sunlight, slow push-in",
+        }}
+      />
+    </Clip>
+  </Render>,
+  { output: "output/hero.mp4", cache: ".cache/ai" },
+);
+```
+
+## Repair Strategy When Results Are Weak
+
+1. Add missing dimensions from the 4D formula.
+2. Increase specificity about motion and camera.
+3. Reduce complexity: fewer subjects, simpler scene.
+4. Lock consistency: reuse `Image(...)` outputs across clips.