diff --git a/.gitignore b/.gitignore
index 27e8dfb0..000826d6 100644
--- a/.gitignore
+++ b/.gitignore
@@ -51,6 +51,7 @@ tests/skill_engine/*
 !tests/skill_engine/test_evolver_length_recovery.py
 !tests/skill_engine/test_evolution_retry_idempotency.py
 !tests/skill_engine/test_analyzer_length_recovery.py
+!tests/skill_engine/test_detect_patch_type_prose.py
 !tests/skill_engine/decision/
 tests/skill_engine/decision/*
 !tests/skill_engine/decision/test_analysis_adapter.py
diff --git a/openspace/skill_engine/patch.py b/openspace/skill_engine/patch.py
index c0668893..dfa5e703 100644
--- a/openspace/skill_engine/patch.py
+++ b/openspace/skill_engine/patch.py
@@ -482,11 +482,14 @@ def detect_patch_type(content: str) -> PatchType:
     if "*** Begin Files" in content:
         return PatchType.FULL
 
-    # Detect bare *** File: markers (no *** Begin Files envelope).
-    # Must appear at line start; require at least one to look structural.
-    file_header_hits = _FILE_HEADER_RE.findall(content)
-    if file_header_hits:
+    # ≥2 File: markers, or one at content start with a body (docstring rule 3).
+    file_headers = list(_FILE_HEADER_RE.finditer(content))
+    if len(file_headers) >= 2:
         return PatchType.FULL
+    if len(file_headers) == 1:
+        match = file_headers[0]
+        if not content[: match.start()].strip() and content[match.end() :].strip():
+            return PatchType.FULL
 
     if "<<<<<<< SEARCH" in content:
         return PatchType.DIFF
@@ -527,6 +530,10 @@ def parse_multi_file_full(content: str) -> Dict[str, str]:
         # No multi-file markers — treat entire content as SKILL.md
         return {SKILL_FILENAME: content}
 
+    # Single mid-document *** File: is prose/example text, not multi-file.
+    if len(headers) == 1 and stripped[: headers[0].start()].strip():
+        return {SKILL_FILENAME: content}
+
     files: Dict[str, str] = {}
     for i, match in enumerate(headers):
         file_path = match.group(1).strip()
diff --git a/tests/skill_engine/test_detect_patch_type_prose.py b/tests/skill_engine/test_detect_patch_type_prose.py
new file mode 100644
index 00000000..98c04fe8
--- /dev/null
+++ b/tests/skill_engine/test_detect_patch_type_prose.py
@@ -0,0 +1,38 @@
+"""Regression: mid-document *** File: prose must not drop SKILL.md."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from openspace.skill_engine.patch import (
+    PatchType,
+    create_skill,
+    detect_patch_type,
+    parse_multi_file_full,
+)
+
+
+def test_detect_patch_type_ignores_single_mid_document_file_marker():
+    content = "# My Skill\n\n*** File: helper.sh\n#!/bin/bash\necho hi\n"
+    assert detect_patch_type(content) == PatchType.FULL
+    parsed = parse_multi_file_full(content)
+    assert list(parsed.keys()) == ["SKILL.md"]
+    assert "*** File: helper.sh" in parsed["SKILL.md"]
+
+
+def test_create_skill_keeps_skill_md_when_file_marker_is_prose(tmp_path: Path):
+    content = "# My Skill\n\n*** File: helper.sh\n#!/bin/bash\necho hi\n"
+    result = create_skill(tmp_path / "s", content)
+    assert result.ok
+    skill_md = tmp_path / "s" / "SKILL.md"
+    assert skill_md.is_file()
+    assert "*** File: helper.sh" in skill_md.read_text()
+    assert not (tmp_path / "s" / "helper.sh").exists()
+
+
+def test_detect_patch_type_accepts_structural_multi_file_at_start():
+    content = "*** File: SKILL.md\n# Skill\n\n*** File: helper.sh\n#!/bin/bash\necho hi\n"
+    assert detect_patch_type(content) == PatchType.FULL
+    parsed = parse_multi_file_full(content)
+    assert "SKILL.md" in parsed
+    assert "helper.sh" in parsed