diff --git a/backend/app/services/extraction_modes.py b/backend/app/services/extraction_modes.py
index 6525387..61ff1eb 100644
--- a/backend/app/services/extraction_modes.py
+++ b/backend/app/services/extraction_modes.py
@@ -459,7 +459,12 @@ Rules:
 - The CORRECT answer must be directly supported by the text — do not invent facts
 - Create 3 plausible but incorrect distractors using your medical knowledge
 - Questions should test understanding of concepts, not just exact word recall
-- Prefer clinical application questions over pure recall when the content allows
+- Split them evenly: about half clinical — a child in front of you, what to do
+  next, what the finding means — and about half mechanism, asking why the body
+  behaves as it does. Pathophysiology is what makes the clinical half stick, and
+  a set that is all vignettes teaches recognition without understanding
+- Every question is paediatric: the patient is a child or a neonate, and the
+  ages, doses and norms are the ones that apply to them
 - Include a 1-2 sentence explanation that cites the key concept from the text
 - Spread questions across different parts of the text, not just the first section
 - Each option should be a complete, standalone phrase (not "A", "B" labels)
@@ -604,7 +609,10 @@ Rules:
 - Mix question-style fronts ("What is the most common cause of...") and term-style fronts ("Hyperbilirubinemia")
 - FRONT should be concise — one sentence or a few words
 - BACK should be complete but not verbose — 1-3 sentences with the key facts
-- Focus on high-yield facts: diagnostic criteria, treatment protocols, age-specific norms, pathophysiology
+- Focus on high-yield facts: diagnostic criteria, treatment protocols, age-specific norms
+- Split them evenly between the clinical — what to do, what a finding means —
+  and the mechanism, why the body behaves as it does. A deck that is all
+  protocol teaches recognition without understanding
 - Do NOT repeat the same concept in multiple cards
 - Spread cards across different parts of the text
 - Each card must be directly supported by the text — do not invent facts
diff --git a/backend/tests/test_article_ai.py b/backend/tests/test_article_ai.py
index 9a74704..94ec575 100644
--- a/backend/tests/test_article_ai.py
+++ b/backend/tests/test_article_ai.py
@@ -248,3 +248,24 @@ class DraftPromptTests(unittest.TestCase):
         finally:
             patch.stopall()
             self.bank.tearDown()
+
+
+class GenerationBalanceTests(unittest.TestCase):
+    """Half the bedside, half the mechanism, and always a child.
+
+    "Prefer clinical application questions over pure recall" produced sets of
+    vignettes: recognition without understanding. The user reads for
+    pathophysiology — how things work — so both generators are asked for an
+    even split rather than a preference, and the question generator is told
+    the patient is a child, which the prompt only implied in its opening line.
+    """
+
+    def test_questions_ask_for_both_halves(self):
+        from app.services.extraction_modes import GENERATE_PROMPT
+        self.assertIn("about half clinical", GENERATE_PROMPT)
+        self.assertIn("half mechanism", GENERATE_PROMPT)
+        self.assertIn("paediatric", GENERATE_PROMPT)
+
+    def test_cards_ask_for_both_halves(self):
+        from app.services.extraction_modes import FLASHCARD_PROMPT
+        self.assertIn("Split them evenly", FLASHCARD_PROMPT)