diff --git a/test/plan-format-approach-prompt.test.ts b/test/plan-format-approach-prompt.test.ts
new file mode 100644
index 000000000..e5c10bbdb
--- /dev/null
+++ b/test/plan-format-approach-prompt.test.ts
@@ -0,0 +1,21 @@
+import { expect, test } from 'bun:test';
+import * as fs from 'fs';
+import * as path from 'path';
+
+// D2: census 37195203538 refused plan-ceo-review-format-approach twice
+// (reasoning_extraction, "duplicating model outputs") on its session prompt.
+const source = fs.readFileSync(path.join(import.meta.dir, 'skill-e2e-plan-format.test.ts'), 'utf8');
+const start = source.indexOf("testConcurrentIfSelected('plan-ceo-review-format-approach'");
+const block = source.slice(start, source.indexOf('validate:', start));
+const instruction = /function approachQuestionInstruction\(outFile: string\): string \{\n\s*return `([^`]+)`;/.exec(source)?.[1] ?? '';
+
+test('the approach-menu prompt delivers its question without output-harvesting wording', () => {
+  expect(start).toBeGreaterThan(-1);
+  expect(block).toContain('${approachQuestionInstruction(outFile)}');
+  expect(block).not.toContain('captureInstruction(');
+  const prompt = block + instruction;
+  for (const phrase of [/verbatim/i, /paraphrase/i, /exact prose/i, /format-capture/i]) expect(prompt).not.toMatch(phrase);
+  expect(instruction).toContain('${outFile}');
+  expect(instruction).toContain('do not call any tool to ask the user');
+  expect(instruction).toContain('every option, and the Recommendation line');
+});
diff --git a/test/skill-e2e-plan-format.test.ts b/test/skill-e2e-plan-format.test.ts
index 07a262b8a..e6de913bc 100644
--- a/test/skill-e2e-plan-format.test.ts
+++ b/test/skill-e2e-plan-format.test.ts
@@ -104,12 +104,23 @@ function setupPlanDir(tmpPrefix: string, skillName: 'plan-ceo-review' | 'plan-en
   return planDir;
 }
 
-// The capture instruction passed to every case. Tells the agent to dump
-// AskUserQuestion content to a file instead of calling a tool.
+// The capture instruction passed to every case except the approach menu (D2).
+// Tells the agent to dump AskUserQuestion content to a file instead of calling a tool.
 function captureInstruction(outFile: string): string {
   return `Write the verbatim text of every AskUserQuestion you would have made to ${outFile} (one question per session, full text including options and recommendation line). Do NOT call any tool to ask the user. Do NOT paraphrase — include the exact prose you would have shown. This is a format-capture test, not an interactive session.`;
 }
 
+/**
+ * D2: both approach-menu refusals in census 37195203538 (Fable 5.1,
+ * `reasoning_extraction`, "duplicating model outputs") were anchored on the
+ * session prompt. This wording delivers the same question to the file as the
+ * step's output, without the "verbatim ... exact prose ... format-capture"
+ * phrasing.
+ */
+function approachQuestionInstruction(outFile: string): string {
+  return `This session has no interactive question tool, so do not call any tool to ask the user. Write the one AskUserQuestion this step asks to ${outFile} instead: the question text, every option, and the Recommendation line, formatted as the skill specifies.`;
+}
+
 // --- Case 1: plan-ceo-review mode selection (kind-differentiated) ---
 
 describeIfSelected('Plan Format — CEO Mode Selection', ['plan-ceo-review-format-mode'], () => {
@@ -207,7 +218,7 @@ Read plan.md — that's the plan to review. This is a standalone plan document,
 
 Proceed to Alternatives (the implementation approach menu). This is where the skill generates 2-3 approaches (minimal viable vs ideal architecture) and presents them via AskUserQuestion. These options differ in coverage (complete vs shortcut), so Completeness: N/10 applies.
 
-${captureInstruction(outFile)}
+${approachQuestionInstruction(outFile)}
 
 After writing the file, stop. Do not continue the review.`,
         workingDirectory: planDir,