A style evaluation skill for mizchi's personal blog. After writing an article, have a subagent judge "mizchi style-ness" and "AI smell-ness" and iteratively improve.
For the detailed style profile (a collection of 5 characteristic phrases + structural analysis of 23 past articles), see ~/ghq/github.com/mizchi/zenn/CLAUDE.md (644 lines). This skill focuses on an excerpt of its key points + AI-smell criteria + the evaluation workflow.
When to use
- Immediately after writing a draft for a zenn / dev.to / personal blog article
- When you want to bring an AI-drafted article closer to mizchi's style
- When it "smells like AI" but you can't articulate specifically where
When not to use:
- Business documents / OSS READMEs (publicly in English, technical reference tone)
- Pair-programming chat responses (real-time conversation)
The core of mizchi's style (5 axes)
Details are in zenn/CLAUDE.md, but the checkpoints for review can be consolidated into these 5 axes:
- tl;dr up front: 3-5 bullet points at the start of the article. Lead with the conclusion to draw readers in
- Explicit subjectivity: Make the judging subject clear, e.g. "from my experience," "I think," "my conclusion." Don't pretend to be neutral
- Colloquial mixing: Mix spoken language into technical articles, e.g. "めっちゃ" (super), "シュッと" (smoothly), "だるい" (a pain), "消耗した" (drained), "〜っぽい" (-ish), "なんか" (kinda). Has the effect of closing the distance
- Concreteness and code: Always include concrete commands / code / numbers / screenshots between abstract discussions. Factual reports like "it worked"
- Frank disclosure of limits: Clearly state what you can't do and the challenges, e.g. "needs a bit more research," "ended up being a compromise," "use at your own risk"
Sentence endings are often "〜と思う" (I think), "〜だろう" (probably), "〜のはず" (should be). Avoid assertions while making subjectivity clear.
Typical AI-smell patterns (to avoid)
Signals to check in review for "this makes it smell like AI":
| Pattern | AI example | mizchi fix |
|---|---|---|
| Verbose polite expressions | "〜することができます" (it is possible to), "〜することが重要です" (it is important to) | "〜できる" (can), "〜が大事" (matters) |
| Empty emphasis words | "素晴らしい" (wonderful), "画期的な" (groundbreaking), "革新的な" (innovative), "重要な" (important) | Delete, or replace with something concrete |
| Meaningless neutrality that presents both sides | "〜という意見もあれば〜という見方もある" (some say X, others see it as Y) | Make your own position clear on which it is |
| Summary of the previous chapter at the start of a chapter | "In the previous chapter, I discussed X. In this chapter..." | Delete |
| Closing with generalities | "The important thing is balance" / "Let's use them appropriately" | Replace with concrete judgment via "In my experience, ..." |
| Symmetrical heading hierarchy | "Overview / Pros / Cons / Summary" in every section | Make each section's length asymmetrical according to its content |
| Absence of subjectivity | "It is generally said that ..." | "I think that ..." |
| Excessive bullet points | Bullet everything that could be written in prose | Claims in sentences, bullets only where enumeration is needed |
| Repetition in the conclusion | "That concludes this article's explanation of ..." | End with "The end" or the final claim |
| Generic examples prone to hallucination | "For example, one company has an initiative called ..." | Replace with your own repo / actual commits / real numbers |
Evaluation workflow (adapted from empirical-prompt-tuning)
empirical-prompt-tuning The skill's workflow is mostly reused, but the target is not scenarios but "article files."
- Baseline: Write the article draft.
zenn/CLAUDE.mdBe conscious of which of the composition patterns (the 8 types: technical explanation / problem solving / design policy / new technology evaluation / idea proposal / implementation tutorial / manifesto / experiment report) it falls under - subagent dispatch: Launch a new subagent with the Task tool and request a review of "mizchi style score" and "AI smell score" using the template described below
- Two-sided evaluation: Extract the following from the subagent's report
- mizchi score (0-10): Score the 5 axes of tl;dr / subjectivity / colloquialism / concreteness / frankness at 0-2 points each, then total
- AI smell score (0-10): Deduct 1 point per hit for the 10 patterns above (10 - hits, minimum 0)
- Flagged locations: A list with line numbers of what the subagent cited as "this doesn't feel like mizchi / this smells like AI"
- Applying diffs: Fix the flagged locations that you can agree with. You don't have to eliminate all of them in 1 iter (up to 3-5 related locations)
- Re-evaluation: Re-review with a new subagent
- Convergence check: Stop when "mizchi score ≥ 8 and AI smell score ≥ 8 for 2 consecutive iters" or when "the quality of the feedback reaches the level of 'preferences only the person themself can judge'"
Don't aim for perfection.Shipping at 8 / 10 is the right answer. Style is an area where human judgment should remain.
subagent dispatch template
あなたは mizchi (https://zenn.dev/mizchi) の長年の読者です。
今回、新しい記事ドラフトをレビューしてほしい。
## 対象記事
<記事ファイルのパス>
## 文体プロファイル
mizchi 文体の核を理解するため、まず次を Read してください:
- /Users/mz/.claude/skills/mizchi-blog-style/SKILL.md (本 skill 本体)
- /Users/mz/ghq/github.com/mizchi/zenn/CLAUDE.md (詳細プロファイル 644 行、特に "特徴的フレーズ集" "AI 臭の典型パターン" 節)
## レビュー観点
### A. mizchi 度 (0〜10)
次の 5 軸を各 0〜2 点で採点、合計。
1. tl;dr 先出し: 冒頭で結論を 3〜5 箇条で出している
2. 主観の明示: 「自分は」「自分が思う」等で判断主体を明確化
3. 口語混入: 「めっちゃ」「シュッと」等の口語が適度に混じる (過剰でも不足でもなく)
4. 具体性: 抽象論ばかりでなく具体コマンド / コード / 数値 / 実例がある
5. 率直な限界開示: 課題 / 失敗 / 妥協を隠さず書いている
### B. AI 臭度 (0〜10)
SKILL.md「AI 臭の典型パターン」表の 10 項目について、本文で該当する箇所を行番号 + 引用付きでリストアップ。
スコアは「10 - 該当件数」(最低 0)。
### C. 個別指摘
特に強い違和感を覚えた箇所 (mizchi っぽくない / AI 臭い) を最大 5 件、行番号 + 引用 + 改善案 で書く。
## レポート構造
- mizchi 度: X/10 (内訳 1: x, 2: x, 3: x, 4: x, 5: x)
- AI 臭度: X/10 (該当パターン: ...)
- 個別指摘 Top 5: <行番号> "<引用>" → <改善案>
- 総評: 1 段落
Pitfalls
- Leaving it to the subagent to "rewrite it in mizchi style": AI can't fully eliminate AI smell. Stop at the judgment stage, and do the fixes by hand yourself (or via the parent agent)
- Aiming for mizchi 10/10: Overusing "めっちゃ" and "シュッと" makes it unnatural in the opposite direction (parody). 8 / 10 is enough
- Chasing only AI-smell detection: Without adding the 3 axes of subjectivity / concreteness / frankness, it just becomes a "cold technical document"
- Reusing the same subagent: It learns from the previous feedback; dispatch a new one each time
- Have it read the style skill before writing the draft: It over-adapts. Keep the order of writing first, then having it review
Related
~/ghq/github.com/mizchi/zenn/CLAUDE.md— Detailed style profile (main body)empirical-prompt-tuning— Meta skill for subagent dispatch + iterative improvement (the workflow foundation of this skill)