# "Doesn't read right": a Japanese engineer with 3,500 followers got 2,857 stars from one tweet about an anti-AI-tone skill · GitHub open-source project

> Source: https://stringers.io/en/github/gh_nanaism__yomiyasu.html (Stringers / 什么值得做, generated 2026-11-08)
> Stringers answers one question every day: what small product can one or two people build in a few months and actually make money from. Numbers come from 15 databases we collect daily, each with source and date.
> For agents: these verdicts and the underlying databases (verified revenue, App Store / Google Play charts, ad libraries, GitHub, Product Hunt…) are queryable live over MCP — tools today / hot / products / directions / direction / data. Setup: https://github.com/komako-workshop/stringers (the user logs in at stringers.io → "配置 agent" to get a personal URL). Site index: https://stringers.io/llms.txt

> Japanese engineers were fed up with AI-sounding writing. The author posted a tweet with the line "banned words don't work, change the structure" and a few AI-sounding examples everyone had seen. It got 3.2M views, then landed on Hatena's hot entries and Qiita's weekly top. Near-daily releases after that kept the buzz going for a month.

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-6

nanaism/yomiyasu · nanaism

# "Doesn't Waste Your Time" reads awkwardly: a 3,500-fan Japanese engineer's skill, boosted by one tweet, hits 2,857 stars

**What it is** A skill you install into Claude Code, Codex, or Cursor that rewrites awkward AI-written Japanese into natural prose: it restores who did what to whom, swaps metaphorical verbs for concrete ones, and ships with test corpora and a check script. Free and open source.

**Why it took off** A Japanese engineer fed up with AI-speak wrote a tweet arguing that banning words doesn't work, you have to change the structure, with a few AI-speak examples everyone has seen. It got 3.2M views, then hit the Hatena hot list and Qiita's weekly chart. Near-daily releases kept the momentum going for a month.

**B** **How you could follow it** Languages like German, Spanish, Portuguese, and French, which only have humanizer translations, can use the same approach: document the syntactic habits of AI-written text in your language, add a corpus and check script, and open the launch post with a before-and-after that makes local readers laugh.

Grade B: the playbook is clear, and a small account made it work, but Chinese, Korean, and English already have projects with tens of thousands of stars, and the remaining languages have a low ceiling. Whether one of them takes off depends on luck with a single tweet.

Data (updated daily) · Repo created 2026-09-30 · 2,896 stars · 7 days +122

GitHub · stars

Data 2026-10-01 → 2026-11-08 (39 days, 32 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

yomiyasu (よみやす, "easy to read") is an Agent Skill built by a 25-year-old engineer in Fukuoka. Install it into Claude Code, Codex, or Cursor and it rewrites awkward AI-generated Japanese. Anti-AI-speak skills already existed, and a Japanese-specific predecessor had two thousand stars, so yomiyasu came late. Still, it reached 37 stars in 2,857 days (GitHub, 11-06), and its install count on skills.sh passed the 65k-star English original. What set it apart was one line nobody else had said, and a tweet seen by 3.2M people.

### A four-line tweet

On Sep 30, 2026, at nearly 12 a.m. Japan time, Aiichiro Oga (X @oga_aiichiro) posted a tweet. He's an engineer at ALGO ARTIS, an optimization startup, and had about 3,500 followers on X at the time. The tweet was four lines:

> I released a skill called 'yomiyasu' that rewrites awkward, hard-to-read AI-written text into natural, readable Japanese. Improved over the old 'banned-word' approach, it focuses on restoring syntactic structure (who did what to what) and making metaphorical verbs concrete.

In plain terms: the old approach banned certain words; this one changes sentence structure, restoring "who did what to what" and replacing figurative verbs with plain ones.

The tweet ended with 9,104 likes, 911 reposts, 14k bookmarks, and 3.2M views (X, 11-06). Bookmarks outnumbered reposts more than tenfold, meaning most people reacted with "save this, I'll need it soon." X even auto-generated a trending summary for it. A few hours later the author published a long post on his company's Zenn blog. In Oct 1, that article and the GitHub repo both reached Hatena Bookmark's hot list. The article alone got 357 bookmarks that day, and the repo link was shared 204 times.

### What kind of Japanese it fixes

The long post opens with a typical AI-written technical passage: data "quietly breaks" (静かに壊れます), a principle "works in a low-key way" (地味に効いてくる), and you shouldn't let debugging "melt your time" (時間を溶かさない). Japanese engineers read that and smile in recognition. The meaning sort of comes through, but you can't tell who did what or where the problem is.

The author's argument is that banning words doesn't work. Ban "地味に効く" in a prompt and the AI swaps in "効果を発揮する"; the skeleton of the sentence stays the same and still reads awkwardly. Ask it to be "more human" and it piles on "真理", "境地", and "美学". So yomiyasu doesn't use a blocklist. It changes sentence structure: it aligns subjects and predicates, drops the habit of giving tools human-like personality, and replaces verbs like 壊れる / 倒す / 効く / 溶かす with plain wording.

A SKILL.md alone wouldn't look different from other skills. yomiyasu also includes:

- **160 test corpus texts** that specifically distinguish normal usage like 氷を溶かす from the AI-specific 時間を溶かす.

- **A Python check script** that automatically compares before and after, using only the standard library. AI tends to go easy on grading its own work, so the checking is left to a script.

- **An A/B testing app** the author uses to repeatedly compare where two versions read better and where they feel awkward.

One line, `npx skills add nanaism/yomiyasu`, installs it, and telling your AI "use yomiyasu to revise this" triggers it. Open source, MIT licensed, free.

### Where the stars came from

| Date | What happened | star |
|---|---|---|
| 9-30 night | Launch tweet, 3.2M views | — |
| 10-01 | Zenn article and repo hit Hatena's hot list; @tetumemo, with 52k followers, wrote a long review, 210k views | +195 |
| 10-02 | Reached #8 on GitHub Trending's Python chart | +239, the highest single-day gain of the run |
| 10-03 to 10-07 | A Qiita post titled "yomiyasu Went Viral, So I Compared Three" landed #4 on Qiita's official weekly chart | About 100 a day |
| 10-16 / 10-17 | v2.0.0 and v2.0.1 shipped, tweet 130k views | 10-17 +125 |
| 10-24 | Three versions shipped together (v2.3.0, mini, pocket), tweet 400k views | +55 |
| November | No more big moments | 12 to 20 a day |

(Day-by-day October star counts come from a third-party tracker; figures after 10-17 come from GitHub's daily data. View counts are from X, 11-06.)

@tetumemo's long review argued that at first glance it looks like the other Humanizer-style tools, but after reading the author's write-up the approach is different. Other tools "find the words and rhythms AI likes and rewrite them"; yomiyasu "takes the sentence skeleton apart and rebuilds it." He also compared it one by one against five existing Japanese tools in the same category. Once a big account spelled out the difference, readers remembered it. Hatena's comments were split, with some saying "I want to try it" and others calling it "junk from the AI era," and the argument brought in another wave of attention.

After that, the author shipped almost daily, releasing 26 versions in 37 days. Each major version got a tweet and an update to the Zenn article's title (now "Latest v2.4.0"). He handled one incident well: after v1 shipped, someone pointed out that the rewrite changed the meaning. In v2.0.0 he made "don't change the original meaning, don't add information the source lacks" the first rule, and the article admitted that the README examples had added content the original didn't contain. The admission itself became another release tweet.

Two simplified versions followed: mini for environments with SKILL.md size limits, and pocket, only 400 to 500 characters, which can be pasted straight into ChatGPT custom instructions. People who don't use coding tools could use it too.

### Who's nearby

| Project | Scale (11-06) | Compared with yomiyasu |
|---|---|---|
| blader/humanizer | 65k stars; 12.1k installs on skills.sh | The English original, revises according to a checklist of AI-writing traits |
| coji/natural-japanese | 2,024 stars; 5,200 installs on skills.sh | Japanese, two and a half months earlier; rewrites more heavily, rearranging paragraphs |
| textlint-rule-preset-ai-writing | 1,159 star | Rule-based checking that only reports problems and doesn't rewrite |
| iKora128/stop-ai-slop-jp | 477 star | A Japanese review checklist, focused on vocabulary and exaggerated headlines |
| gonta223/humanizer-ja | 163 star | A Japanese version of humanizer, still the blocklist approach |
| bamboo-nova/meiseki | 46 star | Similar idea, but nobody promoted it |

yomiyasu has 17k installs on skills.sh, more than all of the above. A Qiita comparison review found that it "keeps the most of the original," and that natural-japanese has its own strengths. The product is good, but it also got described first as the category's defining example, so later projects are measured against it.

Copycats have already appeared: a Go port, versions styled as persistent Claude Code behavior, an MCP version, and hooks that intercept output. Each has only 0 to 5 stars. Building another yomiyasu in Japanese wouldn't add much.

### Why this one

- **The mood lined up** : the same week, Qiita's weekly #1 was "tired of AI-generated slides," and #8 was "don't want to read AI-written docs." Japanese engineers were fed up with AI-speak at the time.

- **One line explained the difference** : "banning words doesn't work, you have to change the structure." That quietly labeled every Humanizer-style tool as a band-aid.

- **例子好笑** :「时间を溶かさない」这种句子不用解释,每个人都在自己的 PR 里见过。

- **It looked rigorous** : corpora, scripts, A/B testing, and citations are far more credible in tech circles than "yet another prompt."

- **Japan's tech distribution chain is very short** : X → Hatena → Qiita / Zenn / note. A single small account's post can travel the whole path in one day.

The product differentiation is real, but it wasn't enough by itself: natural-japanese was of decent quality but never had that tweet with 3.2M views. The author had few followers, which suggests the momentum came from the tweet itself, not from his reputation. (Inference: with a less funny example, it might not have taken off.)

### If it were you

**Verdict** : B. The playbook of picking one language, fighting only AI-speak, pitching an anti-blocklist approach, and shipping a corpus and check script has worked for someone with 3,500 followers, so it can be copied. But the biggest languages are already crowded, the rest have a low ceiling, and whether one takes off depends on luck with a single tweet.

**See who's already there first** (GitHub, 11-06): Chinese has op7418/Humanizer-zh at 22k stars and lieflat-less-ai-tone at 6,300 stars; Korean has epoko77-ai/im-not-ai at 6,400 stars; Russian's humanizer-ru has 500 stars. German (humanizer-de 197), Brazilian Portuguese (humanizer-pt-br 64), and French (humanizer-fr 20) still only have translated humanizer versions built on the blocklist approach yomiyasu used as its foil. Spanish has almost nothing.

**Who it suits** : developers whose first language is German, Spanish, Portuguese, or French, who can see exactly where AI-speak goes wrong in their language, and who have a local developer community (X, Reddit, local tech blogging platforms) where long posts can run.

**Who it doesn't suit** : anyone trying to build another one in Chinese, Korean, or English, where five-figure-star projects already exist. Also not for people who can't articulate "what exactly is weird about AI-speak in our language." That is a matter of intuition, and you can't translate someone else's rules.

**How to do it** (pick one):

1. **Change the language, keep the method** . Don't translate humanizer's banned-word list. Instead, document the syntactic habits of AI-written text in your language (you'll have to find out what they are for Spanish or Portuguese), add dozens to hundreds of test texts and a working check script, and release a short version that can be pasted into ChatGPT's custom instructions.

2. **Cover a use case yomiyasu skipped** . It only handles technical documentation. AI-speak is just as heavy in customer support replies, marketing copy, and academic papers. academic-humanizer already has 3,600 stars, which shows people will want niche versions.

**How to get users** : don't open the launch post with "another anti-AI-speak tool." Lead with "banning words doesn't work," then post a sample of AI-speak that local readers will recognize as funny, alongside the revised version. Also publish a long post with data and citations on local tech communities. After each major release, post an update, and when someone finds a mistake, fix it publicly and admit it.

**When to act** : the sooner the better. Translated humanizer versions are appearing in each language one by one, and the models' own writing keeps improving. Once a model's default style stops having this tone, demand for skills like this will drop.

**Signs you misjudged it** : fewer than 200 stars two weeks after launch, or replies saying "the model I use doesn't write like that anymore."

[GitHub repo ↗](https://github.com/nanaism/yomiyasu)

**Research notes: competitor table, acquisition evidence, full research report**

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| blader/humanizer | Open source | 2026-01-18 | 65.5K stars; 12.1k installs on skills.sh | English-first, scrubbed of AI-writing tells per a checklist; yomiyasu focuses on Japanese syntax, with test corpora and a check script |
| coji/natural-japanese | Open source | 2026-07-12 | 2,024 stars; 5.2K installs on skills.sh | Aimed at polished working Japanese; it restructures paragraphs and rewrites heavily. yomiyasu preserves the original meaning and changes as little as possible |
| textlint-rule-preset-ai-writing | Open source | 2025-06-16 | 1,159 star | Rule-based lint that only detects words and phrases, no rewriting |
| iKora128/stop-ai-slop-jp | Open source | 2026-06-09 | 477 star | A review checklist focused on vocabulary and hype phrasing |
| gonta223/humanizer-ja | Open source | 2026-01-27 | 163 star | A Japanese take on humanizer, 20 feature checklist, still a blocklist approach |
| nwiizo/suiko | Open source | 2026-08 | Not verified | A Rust CLI that diagnoses with fixed rules, no rewriting |
| bamboo-nova/meiseki | Open source | 2026-06-28 | 46 star | Similar idea (reducing reader burden), but it never got promoted and didn't take off |
| Claude / ChatGPT style settings and output styles | Big-company features | — | — | Generic settings, nothing specifically targeting AI-sounding Japanese |

Where users come from

- **Verified** **X author's launch tweet** : 2026-09-30 @oga_aiichiro: 9,104 likes, 911 reposts, 14,148 bookmarks, 3.2M views, author has about 3,497 followers (fxtwitter); X generated a Trending summary

- **Verified** **Big account quote-post** : @tetumemo (52k followers) 10-01 long-form take: 1,294 likes, 210k views, about how it differs from Humanizer-style tools

- **Verified** **Hatena Bookmark popular** : Zenn article with 540 bookmarks (10-01 added that day 357), GitHub URL with 325 bookmarks (10-01 added that day 204)

- **Verified** **Zenn long-form (company Publication)** : zenn.dev/algoartis: 700 likes / 240 bookmarks, continuously updated through v2.4.0

- **Verified** **Third-party reviews and UGC** : a Qiita three-way comparison with 104 likes, ranked #4 on Qiita's official weekly chart; multiple note hands-on reviews; YouTube lectures and comparison videos; Spotify podcast

- **Verified** **Skill distribution sites** : 17k installs on skills.sh, more than humanizer's 12.1k and natural-japanese's 5.2K; listed on directory sites like agentskillshub and hotskills

- **Verified** **Frequent releases plus tweets** : 37 days, 26 releases; the v2.0.1 tweet got 130k views, matching 10-17 +125 stars; a three-version same-day tweet got 400k views, matching 10-24 +55

- **Verified** **GitHub Trending** : reached #10-02 on the Python language chart (trendshift) on 8

## nanaism/yomiyasu: why it took off

> Research date 2026-11-06. Channel: GitHub. Data card: repo age 37 days, 2,857 stars (2,871 read via GitHub API that day), 74 forks; last 7 days +157, the prior 7 days +300; single-day max in our data +124 (2026-10-18, since our repo_day tracking only started 10-17).

### One line

An Agent Skill (SKILL.md plus a Python check script) written by a Japanese engineer (Aiichiro Oga, X @oga_aiichiro, at ALGO ARTIS, an optimization startup) that works in Claude Code / Codex / Cursor to turn AI-generated "AI-sounding Japanese" into natural Japanese. Its main move is restructuring sentences: restoring who did what to whom, and swapping metaphorical verbs (壊れる / 倒す / 効く / 溶かす) for plain wording. It doesn't rely on a list of banned words.

### Verdict

**It blew up in Japan's tech circles overnight from a single X tweet, and the product's differentiation fueled the spike.** 2026-09-30 The author's launch tweet got **9,104 likes, 911 reposts, 14k bookmarks, 3.2M views** , while the author himself had only about 3,500 followers (fxtwitter API, read 2026-11-06). The same day, a Zenn long-form and the GitHub repo both hit Hatena Bookmark's popular list: the Zenn article had 540 bookmarks, 357 of them added on 10-01 itself; the GitHub link had 325 bookmarks, 204 added on 10-01. The author then released almost daily (37 days, 26 releases), pairing each major version with a tweet, which produced small star rebounds.

### 1. Competitors: who else is doing this

"Remove the AI taste from AI-written text" skills were already a crowded category by 2026, and several Japanese-specific ones existed before yomiyasu:

| Name | Form | Launched | Scale (2026-11-06) | How yomiyasu differs |
|---|---|---|---|---|
| blader/humanizer | Open-source skill (English) | 2026-01-18 | 65,494 stars; 12.1k installs on skills.sh | English-first, removes traces per an "AI writing traits" checklist (vocabulary, sentence patterns); yomiyasu focuses on Japanese syntax and adds a check script and test corpora |
| coji/natural-japanese | Open-source skill (Japanese) | 2026-07-12 | 2,024 stars; 5.2K installs on skills.sh (2 skills combined) | Aimed at polished working Japanese, rewrites heavily and restructures paragraphs (per a Qiita comparison); yomiyasu prioritizes preserving meaning and changing less |
| textlint-rule-preset-ai-writing | textlint rule set | 2025-06-16 | 1,159 star | Rule-based lint that only detects words and phrases; yomiyasu uses LLM rewriting plus script checks |
| iKora128/stop-ai-slop-jp | Open-source skill (Japanese) | 2026-06-09 | 477 star | Review plus revision checklist, focused on vocabulary and hype headlines |
| gonta223/humanizer-ja | Open-source skill (Japanese) | 2026-01-27 | 163 star | A Japanese version of humanizer, 20 feature checklist, still a blocklist approach |
| nwiizo/suiko | Rust CLI + skill | 2026-08 | Unverified (trendshift lists it as a similar repo) | Fixed rules for diagnosis, reproducible results; no rewriting |
| bamboo-nova/meiseki | Open-source skills | 2026-06-28 | 46 star | Similar idea (reducing reader burden), but nobody promoted it and it didn't take off |
| "Japanese Technical Writing Style Guide" SKILL.md (@kgsi) | A single SKILL.md | 2026 | Has Hatena bookmarks, scale unverified | Just a style guide text, no test corpora or check script |
| Big-company features | Claude / ChatGPT style settings and output styles | — | — | Generic settings, nothing specifically targeting AI-sounding Japanese |

**Finding:** people were building in this category already, and the Japanese-specific natural-japanese had already reached two thousand stars. yomiyasu wasn't first, yet it drew the most attention after launch: **17k installs** on skills.sh, more than humanizer's 12.1k and natural-japanese's 5.2K (skills.sh, 2026-11-06).

### 2. Growth: where users came from

#### Star curve (daily from trendshift, plus our repo_day data)

- 10-01 +195, 10-02 +239 (hit #8 on GitHub Trending Python that day), 10-03 +133, then about 10-07 a day before 100. trendshift's cumulative +2,130 is below the actual 2,857, likely because early data was missed, so read the shape only.

- 10-16 to 10-17 +65 / +125: v2.0.0 and v2.0.1 released, v2.0.1 tweet got 1,494 likes and 130k views. The single-day high in our data, +124, fell on 10-18, the same wave as trendshift's 10-17; the one-day gap is probably because the two sides cut days by different time zones.

- 10-20 +83: around the launch of yomiyasu-pocket.

- 10-24 +55: three new versions shipped together (v2.3.0, mini, pocket), tweet got 1,148 likes and 400k views.

- From November it fell back to about 12 to 20 a day.

#### Channel evidence

1. **X launch tweet (main engine, confirmed)** : 2026-09-30 @oga_aiichiro, 9,104 likes, 14,148 bookmarks, 3.2M views (author about 3,497 followers). X also generated a Trending summary for it, "AI生成日本語の不自然さを解消する「yomiyasu」スキルがリリース" (x.com/i/trending/2105450280469123185,10-01).

2. **Confirmed amplifier repost:** @tetumemo (AI diagrams × fact-checking, 52k followers) 10-01 wrote a long review, 1,294 likes, 210k views. His point was that it "looks a lot like Humanizer-type tools, but after reading the author's article I saw the approach is different," effectively explaining the differentiator to a wider audience for it.

3. **Hatena Bookmark popular (confirmed)** : Zenn article with 540 bookmarks, GitHub page with 325 bookmarks, both concentrated on 10-01. Comments were polarized, with "want to try it" alongside "junk in the AI era," and that debate itself helped it spread.

4. **Zenn long-form (confirmed)** : published under the company's Publication (algoartis), 700 likes, 240 bookmarks. It cites Gyakusenkawa-chan's before-and-after AI statistics on 70k Qiita articles, plus academic literature on translationese, which lends it weight. The article is continuously updated, and its title prefix changed to "latest v2.4.0".

5. **Follow-on content (confirmed)** : a Qiita review titled "yomiyasu がバズっていたので…3つ比べてみた" with 104 likes, ranked #4 on Qiita's official weekly chart (Qiita Zine, 10-07); multiple note posts titled "I used yomiyasu to revise my own articles" (one person rated 801 of their own note articles); YouTube install tutorials (10-17), comparison videos, and a Spotify podcast; on the Chinese side, a trendshift page, and overseas coverage on hermes-ai.net, LinkedIn, and Facebook.

6. **Skill distribution sites (confirmed)** : 17k installs on skills.sh, with upgrades as a one-liner `npx skills update yomiyasu`. Directory sites such as agentskillshub, openskills, hotskills.dev, and proskills.md also list it, adding long-tail traffic.

7. **GitHub Trending (partly confirmed)** : only charted once, on 10-02, reaching #8 on the Python language chart (trendshift). Our data shows trending_days_any=0, because trending only tracks repos at least about 60 days old.

8. **Company backing (confirmed)** : the README and article both carry ALGO ARTIS recruiting promotion. This counts as company tech-brand content, but no paid promotion was found.

### 3. Why it took off (the reasons)

- **It tapped into the strongest mood among Japanese engineers right now** : the same week, Qiita's weekly top #1 was "tired of AI-generated slides", and #8 was "don't want to read docs written by generative AI". yomiyasu landed right in that "sick of AI-sounding writing" mood. Its launch tweet got far more bookmarks (14k) than reposts (911), a sign people felt "I can use this right away".

- **Its pitch set it apart** : "banned-word lists don't work; change the structure (who did what to whom) and the metaphors and verbs." That idea fits in one sentence, and it can be used to push back on existing Humanizer-style tools. The big account @tetumemo picked up exactly this point when quoting it.

- **Before/After examples make it instantly clear** : the opening passage ("静かに壊れます / 地味に効いてくる / 時間を溶かさない") makes Japanese engineers smile in recognition. That's the best material for spreading.

- **It looks rigorous** : 160 test corpora, A/B testing, check scripts, cited references in the article, and in v2.0.1 the checker was optimized from O(n³) to O(n). That makes it look more credible than "yet another prompt", which goes over especially well in the tech crowd.

- **Ongoing iteration, public corrections** : when someone pointed out that a rewrite changed the meaning, the author changed the core principle in v2.0.0 (don't alter the original meaning, don't add information it lacks) and publicly corrected the corpus statistics in the CHANGELOG. Each release got a tweet, which stretched the first wave of buzz into a month-long tail.

- **Low install barrier, works across many tools** : `npx skills add` installs it in one line, and it supports Claude Code, Codex, Cursor, Gemini, and Antigravity. There's also a mini version (for environments with SKILL.md size limits) and a pocket version (400 to 500 characters, pasted straight into ChatGPT custom instructions).

### 4. Differentiation, and whether that was the reason it took off

- **Differentiation** : (1) it moves from a blacklist to syntactic rewriting, targeting subject-predicate relations, inanimate subjects and metaphorical verbs; (2) it puts "don't change the meaning, don't add information" first, so it changes less; (3) it ships with test corpora and a Python check script for automatic before/after comparison; (4) it only handles Japanese, aimed at work documents like technical articles, specs and PR descriptions.

- **Was that the reason** : it's one of the main reasons, but not the only one. The differentiation gave it a pitch people could repeat, and a third-party review confirmed that "yomiyasu preserves the original text the most". But a similar tool, natural-japanese, is also good quality, and reviewers felt each has its uses. yomiyasu reached 17k installs mainly because its launch tweet happened to blow up, followed by Hatena's hot-entries page, which pushed it into being the go-to name in the category (Qiita's headline was literally "Since yomiyasu blew up, I compared three"). Later entrants now use it as the benchmark.

### V. Visibility

**Already widely discussed in Japan's tech circles** : a tweet with 3.2M views, an X Trending summary, Hatena hot entries, a Qiita weekly top review, multiple note posts, YouTube talks. Chinese- and English-language coverage is sporadic (trendshift, hermes-ai.net, LinkedIn). It's open source and free, so there's no public revenue to speak of.

### VI. Copycats

- Derivatives / ports: sudame/yomiyasu-go (a Go port), atimot/yomiyasu-chat and daichi-tomita-vivion/yomiyasu-chat (packaged as always-on Claude Code output styles), studio-yebisu/yomiyasu-studio (a proofreading tool), Komugi-powerplatform/yomiyasu-sem (Stop-hook interception), maybefix/yomiyasu-mcp, murakami-hue/yomiyasu-marketplace. All have 0 to 5 stars.

- Other languages: Chinese "remove AI flavor" skills (e.g. MrGeDiao/shuorenhua, "speak like a human"), and multilingual humanizers (KO/EN/ZH/JA, 385 stars, similar repos on trendshift). Not necessarily copies, but the thinking is related. A Chinese post said "Chinese authors can directly borrow its approach" (excerpt from a trendshift page).

### VII. What to borrow

- **A single-language, single-pain-point skill, paired with a long data-backed article and a tweet with Before/After examples** : in a native-language tech community like Japan's, even a small account can rack up a million views.

- Label the competitor's approach (banned-word lists) as "treating symptoms", and present your own as "treating the cause" (rewriting syntax). One sentence establishes the difference.

- Build the skill as testable engineering: test corpora, check scripts, A/B testing. When users have doubts, you can respond with data, and trust follows.

- Release frequently, tweeting and updating the original article each time to sustain attention; then ship a simplified version to cover environments where the skill can't be installed (the ChatGPT instruction box).

- This playbook can be copied to Chinese, Korean and other languages with serious "AI-sounding" problems. Chinese already has many "remove AI flavor" skills, so competition is tighter there; Korean may still have room (unverified).

### VIII. Unknowns

- The GitHub stargazers endpoint with timestamps returned 404 this time, so I couldn't verify the source of early stars one by one. I could only rely on trendshift's daily data.

- Whether any company or KOL paid for the reposts: no evidence found. @tetumemo's quote looks organic.

- How many of skills.sh's 17k installs are in ongoing use versus installed and forgotten is unknown.

- Our local x-radar corpus didn't capture related tweets; all X data came from fxtwitter and web search.

### Sources

- Repo [github.com/nanaism/yomiyasu](https://github.com/nanaism/yomiyasu); Releases (GitHub API)

- Zenn article [zenn.dev/algoartis/articles/0b1c731881b25c](https://zenn.dev/algoartis/articles/0b1c731881b25c) (Zenn API: 700 likes / 240 bookmarks, published 2026-10-01 03:28 JST)

- Launch tweet [x.com/oga_aiichiro/status/2105309676305072437](https://x.com/oga_aiichiro/status/2105309676305072437); v2.0.1 tweet .../2111279865299361960; tweet announcing all three versions together .../2113812798413398299; @tetumemo quote [x.com/tetumemo/status/2105512869555237012](https://x.com/tetumemo/status/2105512869555237012) (engagement data from api.fxtwitter.com, 2026-11-06)

- X Trending summary [x.com/i/trending/2105450280469123185](https://x.com/i/trending/2105450280469123185)

- Hatena Bookmark jsonlite API (Zenn article / GitHub URL)

- Qiita review [qiita.com/inoyu-qiita/items/0ffe6e74ecaf3aaa8b14](https://qiita.com/inoyu-qiita/items/0ffe6e74ecaf3aaa8b14); Qiita Zine weekly top [qiita.com/official-columns/news/2026-10-07/](https://qiita.com/official-columns/news/2026-10-07/)

- trendshift [trendshift.io/repositories/273372](https://trendshift.io/repositories/273372)

- skills.sh: nanaism/yomiyasu, blader/humanizer, coji/natural-japanese pages

- Competitor star counts: GitHub API, 2026-11-06

- note: note.com/bodybeat, note.com/hacklog_stealth etc.; YouTube PEpmsfmqlkE, AyYKWjn5jos

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.