diff --git a/poseguide/composition_rules.py b/poseguide/composition_rules.py
new file mode 100644
index 0000000..eb518cd
--- /dev/null
+++ b/poseguide/composition_rules.py
@@ -0,0 +1,129 @@
+"""
+Composition rules engine for PoseGuide.
+Implements rule-of-thirds, horizon detection, and headroom analysis.
+"""
+
+from dataclasses import dataclass
+from typing import List, Tuple, Optional
+import math
+
+
+@dataclass
+class CompositionScore:
+    rule_of_thirds: float  # 0-1, how well key points align with thirds grid
+    horizon_level: float   # 0-1, how level the horizon is
+    headroom: float        # 0-1, adequate headroom score
+    overall: float         # weighted composite
+
+
+class CompositionRules:
+    """Analyze photo composition using geometric rules."""
+    
+    GRID_WEIGHT = 0.35
+    HORIZON_WEIGHT = 0.35
+    HEADROOM_WEIGHT = 0.30
+    
+    def __init__(self):
+        self.thirds_x = [1/3, 2/3]
+        self.thirds_y = [1/3, 2/3]
+        self.ideal_headroom_ratio = 0.15  # ideal 15% of frame
+    
+    def evaluate(self, width: int, height: int,
+                 key_points: List[Tuple[float, float]],
+                 horizon_y: Optional[float] = None,
+                 face_center: Optional[Tuple[float, float]] = None) -> CompositionScore:
+        """Score a photo's composition.
+        
+        Args:
+            width, height: Frame dimensions in pixels
+            key_points: List of (x_ratio, y_ratio) key subject points (0-1 normalized)
+            horizon_y: Normalized y-coordinate of horizon line (0=top, 1=bottom)
+            face_center: Normalized (x, y) of subject's face center
+        
+        Returns:
+            CompositionScore with per-rule and overall scores
+        """
+        thirds_score = self._score_rule_of_thirds(key_points)
+        horizon_score = self._score_horizon(horizon_y) if horizon_y is not None else 1.0
+        headroom_score = self._score_headroom(height, face_center)
+        
+        overall = (
+            self.GRID_WEIGHT * thirds_score +
+            self.HORIZON_WEIGHT * horizon_score +
+            self.HEADROOM_WEIGHT * headroom_score
+        )
+        
+        return CompositionScore(
+            rule_of_thirds=round(thirds_score, 3),
+            horizon_level=round(horizon_score, 3),
+            headroom=round(headroom_score, 3),
+            overall=round(overall, 3)
+        )
+    
+    def _score_rule_of_thirds(self, key_points: List[Tuple[float, float]]) -> float:
+        """Score how well key points align with rule-of-thirds grid intersections."""
+        if not key_points:
+            return 0.5  # neutral
+        
+        scores = []
+        for px, py in key_points:
+            # Distance to nearest thirds intersection
+            min_dist = float('inf')
+            for tx in self.thirds_x:
+                for ty in self.thirds_y:
+                    dist = math.sqrt((px - tx) ** 2 + (py - ty) ** 2)
+                    min_dist = min(min_dist, dist)
+            # Convert distance to score (max distance to intersection is ~0.47)
+            scores.append(max(0, 1 - min_dist / 0.47))
+        
+        return sum(scores) / len(scores)
+    
+    def _score_horizon(self, horizon_y: Optional[float]) -> float:
+        """Score horizon levelness and placement."""
+        if horizon_y is None:
+            return 1.0
+        
+        # Penalize tilted horizon (far from 0 or 1)
+        # Ideal horizons are near thirds: 0.33 or 0.66
+        dist_to_third = min(abs(horizon_y - 1/3), abs(horizon_y - 2/3))
+        return max(0, 1 - dist_to_third / 0.33)
+    
+    def _score_headroom(self, height: int, face_center: Optional[Tuple[float, float]]) -> float:
+        """Score adequate headroom for portrait compositions."""
+        if face_center is None:
+            return 1.0  # no face detected, skip
+        
+        _, fy = face_center
+        # Headroom = space above face center to top of frame
+        headroom = fy  # normalized: 0 = top, fy = distance from top
+        
+        # Ideal headroom is ~15% of frame height above face
+        ideal = self.ideal_headroom_ratio
+        deviation = abs(headroom - ideal)
+        
+        return max(0, 1 - deviation / ideal)
+
+
+def analyze_composition(width: int, height: int,
+                        subjects: List[dict]) -> CompositionScore:
+    """Convenience function to analyze composition from subject data.
+    
+    Args:
+        width, height: Frame dimensions
+        subjects: List of dicts with 'key_points', optional 'horizon_y', 'face_center'
+    """
+    engine = CompositionRules()
+    
+    all_points = []
+    horizon_y = None
+    face_center = None
+    
+    for subj in subjects:
+        if "key_points" in subj:
+            all_points.extend(subj["key_points"])
+        if "horizon_y" in subj:
+            horizon_y = subj["horizon_y"]
+        if "face_center" in subj:
+            face_center = subj["face_center"]
+    
+    return engine.evaluate(width, height, all_points, horizon_y, face_center)