diff --git a/scene_camera_adapters.py b/scene_camera_adapters.py
index fdbd56e..d10664c 100644
--- a/scene_camera_adapters.py
+++ b/scene_camera_adapters.py
@@ -1042,10 +1042,11 @@ def scene_camera_directive(
     geometry = pov_camera_geometry_phrase(parsed, compact_labels) if pov_labels else camera_geometry_phrase(parsed, compact_labels)
     geometry_clause = f" ({geometry})" if geometry else ""
     if pov_labels:
+        subject, _pronoun = scene_subject_terms(subject_kind, pov_labels)
         return (
-            f"{profile['layout_label']} from POV{geometry_clause}: {direction_detail}. "
-            f"{distance_detail}; {elevation_detail}; lower foreground is reserved for POV body or hand cues; "
-            f"use the multiangle camera only as first-person spatial geometry."
+            f"{profile['layout_label']} from POV{geometry_clause}: keep {subject} and the action primary; "
+            f"{profile['place']} context stays beside or behind the bodies, not in the lower foreground; "
+            f"lower foreground is reserved for POV body or hand cues; use the multiangle camera only as first-person spatial geometry."
         )
     return (
         f"{profile['layout_label']}{geometry_clause}: {direction_detail}; "
@@ -1098,6 +1099,35 @@ def composition_subject_text(text: str, subject_kind: str) -> str:
     return text
 
 
+ACTION_COMPOSITION_TERMS = (
+    "action geography",
+    "anal",
+    "ass",
+    "body contact",
+    "cowgirl",
+    "doggy",
+    "explicit",
+    "first-person",
+    "foreground hands",
+    "genital",
+    "hips",
+    "kneeling",
+    "mouth",
+    "oral",
+    "penetration",
+    "pov",
+    "rear-entry",
+    "rear entry",
+    "rear-view",
+    "sexual contact",
+)
+
+
+def is_action_specific_composition(text: Any) -> bool:
+    lower = str(text or "").lower()
+    return any(term in lower for term in ACTION_COMPOSITION_TERMS)
+
+
 def contextual_composition_prompt(
     scene_text: Any,
     composition: Any,
@@ -1110,6 +1140,8 @@ def contextual_composition_prompt(
     text = str(composition or "").strip()
     if not text:
         return text
+    if is_action_specific_composition(text):
+        return text
     text = composition_subject_text(text, subject_kind)
     profile = scene_camera_profile(scene_text, scene_entry=scene_entry, theme=theme, profile_key=profile_key)
     if not profile:
diff --git a/tools/prompt_smoke.py b/tools/prompt_smoke.py
index cd18b0e..9555e82 100644
--- a/tools/prompt_smoke.py
+++ b/tools/prompt_smoke.py
@@ -881,6 +881,36 @@ def smoke_row_camera_policy() -> None:
         "coworking lounge frame with the couple near a desk edge" in updated.get("composition", ""),
         "row camera policy did not adapt coworking composition for couple rows",
     )
+    pov_action_row = dict(row)
+    pov_action_row["composition"] = (
+        "first-person rear-view frame looking down at the woman's raised ass, "
+        "with foreground hands and rear-entry contact readable"
+    )
+    pov_action_row["prompt"] = (
+        "A generated adult prompt. Framed as first-person rear-view frame looking down at the woman's raised ass, "
+        "with foreground hands and rear-entry contact readable. Avoid: low quality."
+    )
+    updated_pov_action = row_camera.apply_camera_config(
+        pov_action_row,
+        _orbit_camera(horizontal_angle=180, vertical_angle=-30, zoom=7.5),
+        compact_labels=pb.CAMERA_COMPACT_LABELS,
+    )
+    _expect(
+        "first-person rear-view frame" in updated_pov_action.get("composition", ""),
+        "row camera policy replaced explicit POV action composition with generic location composition",
+    )
+    _expect(
+        "couple's raised ass" not in updated_pov_action.get("composition", ""),
+        "row camera policy rewrote explicit female action anatomy as couple anatomy",
+    )
+    _expect(
+        "coworking lounge frame with the couple near a desk edge" not in updated_pov_action.get("composition", ""),
+        "row camera policy leaked generic coworking composition into explicit POV action frame",
+    )
+    _expect(
+        str(updated_pov_action.get("camera_scene_directive", "")).count(";") <= 3,
+        "POV camera scene directive became too noisy",
+    )
     already_matching_row = dict(row)
     already_matching_row["pov_character_labels"] = []
     already_matching_row["composition"] = "coworking lounge frame with the subjects near a desk edge and tall-window depth behind them"