diff --git a/CHANGELOG.md b/CHANGELOG.md
index 222fdaa..0721255 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -19,6 +19,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   becwright" error — instead of risking a silent misparse. (The `.bec.yaml`
   export bundle was already versioned via `becwright_bec`.)
 
+### Fixed
+- `becwright init --from-claude-md` no longer misreads a per-*function* line count
+  as a per-*file* cap. A phrase like "~50 lines per function, ~800 per file" used
+  to derive `max_lines --max 50` (flagging nearly every file); the file-cap
+  matcher now refuses to bridge across a comma or another number, so an ambiguous
+  soft guideline derives no cap instead of a wrong one. (Surfaced field-testing a
+  real Python repo.)
+
 ### Documentation
 - Documented becwright's **stable contract** in `documentation/usage.md`: the CLI
   exit codes (`0` pass · `1` a blocking rule failed · `2` config/usage problem)
diff --git a/src/becwright/cli.py b/src/becwright/cli.py
index 2be2e65..359d2c3 100644
--- a/src/becwright/cli.py
+++ b/src/becwright/cli.py
@@ -417,10 +417,13 @@ def _read_claude_md(root: Path) -> str | None:
 
 # A line cap tied to *files* (not functions — that needs an AST): "files < 800
 # lines" or "800 lines per file", EN/ES. The file/module anchor avoids mapping a
-# "functions < 50 lines" rule, which becwright cannot enforce.
+# "functions < 50 lines" rule, which becwright cannot enforce. The gap between the
+# number and the file word excludes commas and digits so a match cannot bridge
+# across a clause boundary — e.g. "~50 lines per function, ~800 per file" must not
+# read the function's 50 as the file cap just because "file" appears later.
 _FILE_LINE_CAP = re.compile(
-    r"(?:files?|archivos?|modules?|m[óo]dulos?)\b[^.\n]{0,40}?(\d{2,4})\s*(?:lines?|l[ií]neas?)"
-    r"|(\d{2,4})\s*(?:lines?|l[ií]neas?)\b[^.\n]{0,25}?(?:per\s+|por\s+)?(?:files?|archivos?|modules?|m[óo]dulos?)",
+    r"(?:files?|archivos?|modules?|m[óo]dulos?)\b[^.,\n\d]{0,40}?(\d{2,4})\s*(?:lines?|l[ií]neas?)"
+    r"|(\d{2,4})\s*(?:lines?|l[ií]neas?)\b[^.,\n\d]{0,25}?(?:per\s+|por\s+)?(?:files?|archivos?|modules?|m[óo]dulos?)",
     re.IGNORECASE,
 )
 
diff --git a/tests/test_init.py b/tests/test_init.py
index 85f7ee9..3196f5e 100644
--- a/tests/test_init.py
+++ b/tests/test_init.py
@@ -320,6 +320,16 @@ def test_max_lines_cap_ignores_function_length():
     assert cli._max_lines_cap("Functions must be under 50 lines.") is None
 
 
+def test_max_lines_cap_does_not_bridge_function_cap_to_file_word():
+    # Real CLAUDE.md phrasing (bot-telegram field test): a per-function number and a
+    # per-file number in one clause, comma-joined. The function number (50) must NOT
+    # be captured just because "archivo"/"file" appears later past the comma. The file
+    # number (800) is elliptical (no "líneas" word of its own), so becwright declines
+    # rather than guessing wrong — an ambiguous soft guideline derives no cap.
+    assert cli._max_lines_cap("~50 líneas por función, ~800 por archivo") is None
+    assert cli._max_lines_cap("~50 lines per function, ~800 per file") is None
+
+
 def test_max_lines_cap_out_of_range():
     assert cli._max_lines_cap("keep files under 40 lines") is None      # below 50
     assert cli._max_lines_cap("files must not exceed 9000 lines") is None  # above 5000