diff --git a/spec/numbers.md b/spec/numbers.md
index facaeb1d4..1f6a3694e 100644
--- a/spec/numbers.md
+++ b/spec/numbers.md
@@ -468,3 +468,18 @@
 | 三日間 | 3일간 | 유지 |
 
 사례: `tests/rules/cases/laugh-legs-month-one.tsv`
+
+## NUM-34 수 뒤에 홀로 쓴 コ는 조수사 `개`다
+
+수(아라비아 숫자, 한자 수, 何, 数) 뒤에 가타카나 コ 한 글자만 오면 個를 가타카나로 쓴 조수사라서 정규화 단계가 個로 읽고 `개`다(メール784コっす 메일 784개예요,
+三コください 세 개 주세요, 何コ 몇 개). コ 뒤에 가타카나가 이어지면 그 낱말이다(3コース 3코스, 5コマ, 2コンボ, 100コイン).
+
+| 원문 | 출력 | 구분 |
+|---|---|---|
+| メール784コっす | 메일 784개예요 | 고침 |
+| 201コだよね | 201개지 | 고침 |
+| 三コください | 세 개 주세요 | 고침 |
+| 3コース | 3코스 | 유지 |
+| 100コイン | 100코인 | 유지 |
+
+사례: `tests/rules/cases/ko-counter.tsv`
diff --git a/src/core/pipeline/normalize.cpp b/src/core/pipeline/normalize.cpp
index aff7d53f6..32c1b8d2e 100644
--- a/src/core/pipeline/normalize.cpp
+++ b/src/core/pipeline/normalize.cpp
@@ -673,6 +673,17 @@ std::string NormalizeForms(std::string_view text) {
         } else {
             c = MapForm(c);
         }
+        // A katakana コ alone after a number is the counter 個 written in katakana (784コ 784個, 三コ, 何コ);
+        // before more katakana it starts a word (3コース, 5コマ).
+        if (c == U'コ' && !out.empty()) {
+            const char32_t prev = LastRune(out).cp;
+            const char32_t next = At(text, i).cp;
+            const bool number = (prev >= U'0' && prev <= U'9') || (prev >= U'０' && prev <= U'９') ||
+                                In(prev, U"〇一二三四五六七八九十百千何数");
+            const bool word = (next >= U'ァ' && next <= U'ヺ') || next == U'ー' || (next >= 0xFF66 && next <= 0xFF9F) ||
+                              next == 0x3099 || next == 0x309A;
+            if (number && !word) c = U'個';
+        }
         const text::Rune mark = At(text, i);
         if (const char32_t composed = ComposeKanaPair(c, mark.cp)) {
             c = composed;
diff --git a/src/core/pipeline/normalize.h b/src/core/pipeline/normalize.h
index 1bf03487e..f3170804a 100644
--- a/src/core/pipeline/normalize.h
+++ b/src/core/pipeline/normalize.h
@@ -129,7 +129,8 @@ std::string ComposeKanaPairs(std::string_view text);
 // small vowel or a katakana letter composed (ウ゛ァ ヴァ); the CP932 mapping differences (〜 ～,
 // − －, ‖ ∥, ¢ £ ¬ in full width) and kanji forms outside JIS X 0208 (頰 頬, 𠮟 叱) mapped;
 // ‼ ⁉ ⁈ as two marks (！！ ！？ ？！) without a variation selector after them; U+0000..U+0003 as
-// spaces. Invalid UTF-8 bytes are kept.
+// spaces; a katakana コ alone after a number as the counter 個 (784コ 784個, not 3コース). Invalid UTF-8
+// bytes are kept.
 std::string NormalizeForms(std::string_view text);
 
 // A run of two or more っ or ッ after a letter and before a letter other than a small kana, ー
diff --git a/tests/rules/cases/ko-counter.tsv b/tests/rules/cases/ko-counter.tsv
new file mode 100644
index 000000000..f0abc309c
--- /dev/null
+++ b/tests/rules/cases/ko-counter.tsv
@@ -0,0 +1,17 @@
+# 규칙: NUM-34
+rule	input	expected	kind	layer
+normalize:ko-counter	メール784コっす	has:784개예요	fix	normalize
+normalize:ko-counter	201コだよね	201개지	fix	normalize
+normalize:ko-counter	446コの	446개의	fix	normalize
+normalize:ko-counter	594コ	594개	fix	normalize
+normalize:ko-counter	１２コ	12개	fix	normalize
+normalize:ko-counter	三コください	세 개 주세요	fix	normalize
+normalize:ko-counter	十コある	열 개 있다	fix	normalize
+normalize:ko-counter	何コある？	몇 개 있어?	fix	normalize
+normalize:ko-counter	3コース	3코스	keep	normalize
+normalize:ko-counter	5コマ	lacks:개	keep	normalize
+normalize:ko-counter	2コンボ	2콤보	keep	normalize
+normalize:ko-counter	100コイン	100코인	keep	normalize
+normalize:ko-counter	3コスト	3코스트	keep	normalize
+normalize:ko-counter	3コロ	3코로	keep	normalize
+normalize:ko-counter	コ、コイツ	lacks:개	keep	normalize