diff --git a/home/dot_claude/references/japanese-writing/norms.md b/home/dot_claude/references/japanese-writing/norms.md
index e467134..f4204f6 100644
--- a/home/dot_claude/references/japanese-writing/norms.md
+++ b/home/dot_claude/references/japanese-writing/norms.md
@@ -66,4 +66,11 @@ Detectable post-hoc; the mechanical-lint safety net beneath the structural axis
 - **English-style colon syntax**: `:` placed right after a clause ending in a verb/adjective（❌「実装した: 〜」）→ receive with a noun before the colon, or break into a separate sentence（✅「実装した内容は次のとおり: 〜」）
 - **Translationese vocabulary（翻訳借用語）**: rare kanji compounds that exist only as literal translations of English terms（canonical → 正本・正典、mitigation → 緩和策の直訳連発）and fail the Audience vocabulary test → the settled Japanese term, カタカナ, or a plain paraphrase（「唯一の情報源」）. Each one caught in a real draft goes into the proofreader dictionary (`writing-proofread` skill, `assets/prh.yml`) so the next detection is deterministic
 - **Mannered prose（技巧の誇示）**: a metaphor, personification, or flourish standing in for a direct statement（❌「ログが原因を語ってくれる」「本書は Rust の世界への招待状だ」 ✅「ログを見れば原因がわかる」「本書は Rust の入門書だ」）→ say what you mean; when a literal phrase exists, use it. The reader pays an extra step so the writer can perform
+- **Metaphor verbs and pseudo-concrete words（AI 偏愛語）**: vocabulary that surged in AI-era technical prose, each item standing in for a concrete statement. Judge by one test: would it appear in pre-AI human technical writing? If yes, leave it
+  - Metaphor verbs: 「地味に効いてくる」「（静かに）壊れる」「〜側に倒す」「時間を溶かす」「静かに／黙って」（"silently" の訳）「〜した瞬間」 → plain words of the same breadth: 「役に立つ」「おかしくなる・使えなくなる」「〜を優先する・〜に統一する」「時間を費やした」「気づかないうちに」「〜した時点で」. Literal uses stay（お腹を壊す、植木鉢が倒れる）. Carry over the nuance the metaphor held（「地味に」= small but real、「溶かす」= regret）in plain words. Write 「通知なく」「エラーを出さずに」 only when the source makes clear that no notification or error occurs
+  - Pseudo-concrete words: 「解像度を上げる」「腹落ち」「手触り」「肌感」「温度感」, and 「本質的な」 as an empty modifier → state what was examined or learned, taken from the source only; when the source doesn't say, ask the author rather than invent it
+  - Framing fillers: 「重要なのは〜」「結論から言うと」「〜に他なりません」「〜というわけです」「いかがでしたでしょうか」 → start from the main clause. When 「重要なのは」 carries the actual evaluation, move it into the predicate（「〜が重要です」）instead of deleting it
+  - Not targets: 既定・照合・要点・定番・実測・突き合わせる・入口・土台・担保・核心・潰す（バグを潰す）are long-standing terms in Japanese technical writing — never flag them
 - **Particle omission and padding**: dropped 助詞（「ファイル作成」→「ファイルを作成する」), passive where active reads plainer（「〜される」→「〜する」）, and padding like 「〜することができる」→「〜できる」、「〜を行う」→ the verb itself
+
+Metaphor-verb and pseudo-concrete-word item distilled from nanaism/yomiyasu `references/slop-catalog.md` (MIT, commit f03aecd9c4c4a69e229f3848ecba66c3871f734c), restated in house style — not a verbatim copy.
diff --git a/home/dot_claude/skills/writing-proofread/SKILL.md b/home/dot_claude/skills/writing-proofread/SKILL.md
index 485676f..39e1d42 100644
--- a/home/dot_claude/skills/writing-proofread/SKILL.md
+++ b/home/dot_claude/skills/writing-proofread/SKILL.md
@@ -100,9 +100,9 @@ Add `--genre essay|tech|business` when the genre is clear — it switches to cal
 - If `uv` is unavailable, skip this pass; Pass 4's manual checks are the fallback
 - If `jq` is unavailable, drop `--json` and the redirect and run the script bare — the human-readable format carries the same findings at roughly 3× the size of the jq view
 
-## Pass 3: AI-smell (the 5 categories)
+## Pass 3: AI-smell taxonomy
 
-Read `~/.claude/references/japanese-writing/norms.md` here, once — Pass 4 works from the same read, so don't return to it. Then hunt each of its six categories explicitly: mechanical list templates, hype vocabulary, over-emphasis, English-style colon syntax, translationese vocabulary（翻訳借用語 — words failing the norms' Audience test; unambiguous ones also feed `assets/prh.yml`）, and particle omission / padding (dropped 助詞, passive → active,「することができる」→「できる」等).
+Read `~/.claude/references/japanese-writing/norms.md` here, once — Pass 4 works from the same read, so don't return to it. Then hunt each category of its AI-smell taxonomy explicitly: mechanical list templates, hype vocabulary, over-emphasis, English-style colon syntax, translationese vocabulary（翻訳借用語 — words failing the norms' Audience test; unambiguous ones also feed `assets/prh.yml`）, mannered prose (metaphor or flourish in place of a direct statement), metaphor verbs and pseudo-concrete words（AI 偏愛語 — context-dependent, so judge from the norms rather than add to `assets/prh.yml`）, and particle omission / padding (dropped 助詞, passive → active,「することができる」→「できる」等).
 
 This manual read is also where plain 誤字・誤用・文法ミス get caught — textlint misses many（「とゆう」等）. Report them under Pass 3 in the findings table.