Techniques for Making AI Write Natural Japanese

The grammar is correct, but reading it is tiring. You get sleepy halfway through. You realize "Ah, this was written by AI" in the very first line.

The unique sense of unease when having AI write in Japanese remains a challenge for many people today. Many have experimented and developed skills and systems to resolve this discomfort.

In this article, I summarize practical insights I've verified through daily writing and community achievements, organized from the following four perspectives:

• Model Selection: Why choosing the right model matters more than harness (framework) improvements
• Recommended Writing Procedure: Five steps to reduce reader cognitive load and create sentence rhythm
• Recommended Skills: Essential tools from predecessors that manage sentence length and prevent reader disorientation
• Actual Writing Environment: Repository structure to overcome the wall where "adding rules brings back the AI smell"

First, I want to state clearly that "Model Selection" is more effective than prompts or hooks (harnesses).

As of 2026, Gemini 3.8 Flash is head and shoulders above the rest for Japanese writing, followed by Kimi K3. Claude, once praised for natural prose, has become noticeably stiff and angular due to its specialization for coding. ChatGPT remains mediocre, with overly uniform paragraph lengths and textbook-like structures that leave the most obvious AI traces.

However, relying solely on the model's raw output results in homogeneous sentence rhythms and offloads cognitive burden onto the reader. Procedures and systems are needed to leverage the model's inherent strength.

Recommended Writing Procedure

Integrating current knowledge, the writing procedure we should take now can be summarized into these five steps.

1[Step 1] Information sorting and cutting (discard unnecessary kindness)
2   ↓
3[Step 2] Designing the reader's mental map (global order and connections)
4   ↓
5[Step 3] Designing intonation and variation (intentional heat distribution)
6   ↓
7[Step 4] Separating writing and mechanical inspection (reject via PostToolUse Hook)
8   ↓
9[Step 5] Humans polish the final 10% (record diffs to feed back into rules)

Step 1. Pre-writing information sorting and "cutting"

The biggest reason AI text is unreadable is "offloading cognitive load" (Tsubota-san / @tsubotax). It packs premises, supplements, and preemptive reassurances into single sentences. The iron rule is to cut information not used for "what the other party will do next" before letting the AI write.

• Determine "who is this for, and what decision is needed" in the first two lines
• Assign only one role to each paragraph, such as "option differences," "decision materials," or "verification methods," and discard paragraphs that fit none
• Remove all parenthetical supplements that don't affect decisions if deleted

Step 2. Global design that doesn't break the reader's "mental map"

As pointed out by Keiichiro Shikano (@golden_lucky), text breakdowns arise from "ambushes where intent isn't clear until you read ahead" or "lack of connection to the previous paragraph." Prepare a path so readers don't get lost.

• When introducing a concept for the first time, order it so readers can identify types and roles just by reading sequentially
• Before introducing new information, prepare a path in the reader's mind to receive it
• Always place one sentence connecting to the previous paragraph immediately after a heading

Step 3. Intentionally embedding "bias (variation)"

AI tends to discuss all topics at the same temperature, but the essence of human touch is "bias" (Natsu-san / @art_reflection). Talk obsessively about interesting topics, breeze past premises in one sentence. Intentionally allocate heat intensity during the structural phase.

Step 4. Separating writing and mechanical inspection

The more prohibitions you list in the prompt, the more the model shrinks and shortens sentences. Let it write freely, then mechanically inspect via Hook immediately after file output. On violation, don't just replace words—make it rewrite the entire sentence containing the issue. Separating writing and inspection is key to maintaining rhythm.

Step 5. Humans polish the final 10%

Don't aim for 100 points with AI alone. Leave 90% of the skeleton to AI, and let humans adjust the remaining 10% of breathing room and texture. Record the corrections and feed them back into future prompts or inspection rules.

Recommended Skills

I also introduce essential community-born tools and skills that improve Japanese document quality, which I currently use myself.

natural-japanese (@techtalkjp**

An Agent Skill born from the insight that "AI smell shows in rhythm more than vocabulary" (GitHub Repo / Explanation Article). Based on measured corpora of human and AI text, it mechanically detects variations in sentence length (burstiness), discarding traditional assumptions like "lack of noun endings" or "few sentence-initial repetitions."

https://x.com/techtalkjp/status/2076808889371443679

japanese-tech-writing (@golden_lucky**

A prompt skillified from commercial book editing perspectives (Gist / Blog Post). It eliminates LLM tone that exaggerates stance without adding arguments, such as handle directly, dig deeper, or nothing less than ~, regulating order and subject placement so readers can understand sequentially.

https://x.com/golden_lucky/status/2066751852088336698

Hooks × lint with good patterns (@yugen_matuni**

A system using AI coding tool Hook functions (PostToolUse) to inspect immediately after file writing. For NG expressions, it returns specific correction patterns (good) like "true value lies in ~" alongside, prompting a rewrite of the whole sentence.

https://x.com/yugen_matuni/status/2088251220452679951

My Writing Environment

In my writing repository, I experienced the failure where "the more rules I added, the more sentences shrank and AI smell returned." Adding conventions made AI retreat to safe, short sentences.

From this, I derived the design principle: "Don't reduce rules; separate their locations."

1[Overall Writing Flow]
2
3[Input] style-guide / voice-samples / edit-patterns (judgment criteria only)
4     ↓
5[Writing Execution] writer (Gemini 3.8 Flash)
6     ↓
7[Auto Inspection] PostToolUse Hook (reject via detect-ai-smell.js)
8     ↓
9[Quality Judgment] reviewer (objective evaluation based on rubric.md)
10     ↓
11[Final Polish] Human adjustment & publication (feed diffs back to rules)

1kgsi-note-documents_v2/
2├── CLAUDE.md              # Operational policy, writing modes, agent permissions
3├── GOVERNANCE.md          # Size limits for reference materials, three storage locations
4├── style-guide.md         # Kogiso's writing style (judgment descriptions only, max 4,500 chars)
5├── references/
6│   ├── voice-samples.md   # Style samples (paragraph excerpts by function)
7│   ├── edit-patterns.md   # Ledger of manual corrections (before→after)
8│   └── rubric.md          # Reviewer evaluation criteria (distance benchmarks)
9├── human-bench/           # 5 past human-written articles (writer doesn't read; threshold for machine inspection)
10├── scripts/
11│   ├── detect-ai-smell.js # Machine inspection gate (regex, density check, with 'good' patterns)
12│   ├── clean-japanese-spacing.js # Auto-remove half-width spaces between Japanese/Western chars
13│   └── hooks/             # Scripts for PostToolUse Hooks etc.
14├── editor/                # Local note-style preview & revision editor
15└── note/                  # Article body (final drafts)

This is the screen of my custom local editor (note-editor). Editing on the left reflects instantly to the central note-production-equivalent preview, and partial AI revision of selected ranges can be done from the right pane.

Custom editor. Launched with /note-editor command

Three Storage Locations and Permission Separation

1. Internalization Layer (read by writer): style-guide.md, voice-samples.md. Describes only the writer's judgment criteria thinly (max 4,500 chars).
2. Machine Layer (auto-scanned by Hook): scripts/detect-ai-smell.js. No character limit. Aggregates regex-detectable prohibitions and correction policies (good).
3. Judgment Layer (read by reviewer): references/rubric.md. A separate agent independently evaluates the distance between the finished draft and the goal.

To prevent bloat, when the guide exceeds 80% of its limit, we delete an amount of old rules equal to any new rules added.

Excerpts from Actual Files

Here are core parts excerpted from files actually in operation.

`GOVERNANCE.md` (Principles of location and limit management)

1If materials read by writer and reviewer swell, sentences shrink trying to follow rules, and AI smell returns.
2In this environment, the principle is "rules are not reduced, but relocated," with limits set per location.
3
4[Three Storage Locations]
5- Internalization Layer (read by writer): style-guide.md, voice-samples.md, edit-patterns.md
6  → Keep thin. Only judgment descriptions. Do not checklistify
7- Machine Layer (auto-scanned by Hook): inspections in scripts/detect-ai-smell.js
8  → No limit. All regex-writable prohibitions go here
9- Judgment Layer (read by reviewer): references/rubric.md, references/voice-samples.md
10  → Distance criteria are excerpts. One template per type only when confused

`scripts/detect-ai-smell.js` (Machine inspection and 'good' contrast)

1// detect-ai-smell.js excerpt
2const CHECKS = [
3  {
4    name: "Evasive expressions / Non-committal attitude",
5    patterns: [/一概には言え|メリットもデメリットも|一長一短|状況によります|と言えるでしょう|と考えられます/g],
6    limit: 1,
7    gate: true,
8    good: "Pick a side and state it firmly, placing one sentence of evidence immediately after. Don't close with both sides presented",
9  },
10  {
11    name: "Preview clift 'What's important is...'",
12    patterns: [/重要なのは|大事なのは/g],
13    limit: 1,
14    gate: true,
15    good: "Write the content directly without previewing. 'What's important is X' → 'X is. Because ~'",
16  },
17  {
18    name: "Signboards of abstract contrasts",
19    patterns: [/という構図です|デザイン思考/g],
20    limit: 1,
21    gate: true,
22    good: "Write contrasts as actions of who does what. 'Structure of A and B' → 'A does ~. B does not do ~'",
23  },
24];

`style-guide.md` (Premise instructions to writer)

1This guide is not a checklist. If you ask yourself "Am I violating rules?" while writing, AI smell returns.
2When I write articles, these operate as internalized premises.
3
4Before starting, verbalize the following:
51. Who am I writing as right now (e.g., designer touching AI in practice)
62. What reaction am I writing about (Expectation → Reality → Gap)
73. How will I close (Leave unresolved questions or assert position and wrap up)
84. Intonation design. Which sections to emphasize, which to breeze through in one sentence

Conclusion

No matter how elaborate the harness or excellent the model chosen, if there is no "raw specificity" within the human writer, the text cannot stand.

AI spits out clichés because humans haven't provided material like "what they experienced and what felt wrong." Abstract instructions only yield abstract platitudes.

"When I actually touched it, it was simpler than expected and disappointing"

"I felt a little relieved when the pull request went through late at night"

Only when there are seeds of flesh-and-blood observations, numbers, and proper nouns do Gemini 3.8 Flash's expressive power and Hook-based auto-inspection truly shine.

AI writing is not lazy automation, but an editorial attempt to deliver your own discomfort efficiently to readers. Try starting by reviewing the line you just wrote.

Note: This article itself was written based on the writing environment explained herein.