Introduction

About 1.5 months after launch, “yomiyasu,” a skill that helps with polishing Japanese prose, has received a major update to v3.0.0!

Previously, I compared three Japanese prose-polishing skills, including yomiyasu v1.0.1, by giving them the same AI-sounding passage. At the time, yomiyasu removed some metaphors and bold formatting while preserving the structure and intensity of the text.

I was fortunate to have my previous article quoted on X.

If you want to preserve individuality, yomiyasu; if you want to focus on structure, natural-japanese. Once I knew how to distinguish between them, polishing became easier

However, the previous article summarized the comparison results from v1.0.1.
Because the article was viewed by many people, I left the comparison results as a record from that time so that readers’ impressions would not change depending on when they read it.

Since we’re at it, I’ll add v2.5.0 and v3.0.0 and compare them again!

To begin with the results, in this run v2.5.0 changed metaphors into more concrete expressions, while v3.0.0 showed notable changes that toned down headings and calls to action. It does not seem that the new version automatically fixes everything more aggressively💡 (Interesting...!)

What changed in yomiyasu from v2.5.0 to v3.0.0

v3.0.0 is not simply a shortened version of the instructions used through v2.5.0. The official changelog describes it as “a major version that rebuilt the prose-editing process.”

In v3.0.0, not only were the instructions shortened, but the rules for what to retain and what to remove were also reconsidered.

• Shortened SKILL.md from 46,311 characters to 15,656 characters
• Semantic-preservation rules, edits tailored to the author’s preferences, inspection, and output format are maintained
• Removed some of the detailed rules that were present in v2.5.0
• Added new rules in response to problems found through external comparisons and audits
• In the official validation of 12 pieces, the average cost per piece fell from about $0.85 to $0.59
• For the same 12 pieces, the amount of source text modified increased compared with v2.5.0

The instructions became shorter, and the average cost also fell in the official validation. It has also been reported that, because some rules removed in v2.5.0 are still missing, some passages are not revised in the same way.

The official validation also reported cases where “self-destruction” was replaced with “performance degradation,” weakening the tone, and where the meaning of “noise” was lost from a heading reading “Silence/Noise Section.” The more editing is done for readability, the more room there is for shifts in the original meaning and the writer’s stylistic temperature.

So, using the same rather heavily embellished passage as last time, I examined how the editing changed between v2.5.0 and v3.0.0.

The validation method was the same as last time, with only the version changed

The validation date was November 11, 2026. I used Claude Sonnet via Claude Code, with effort: low, a new session for each run, and --safe-mode to disable other plugins and instructions.

Using the same command as in the previous comparison, I changed only yomiyasu’s SKILL.md, which was passed to --append-system-prompt-file, between v2.5.0 and v3.0.0. I ran each once in a new session.

View the execution command

For v2.5.0, I used the skill I had pinned locally, and obtained v3.0.0 from the tag in the official GitHub repository.

This is the same request as last time.

Not a single character of the original text has been changed.

Have you had enough of writing the same instructions over and over every day?

Every time you write an article, have you asked AI things like “use desu/masu style,” “make it beginner-friendly,” and “use this heading format” all over again? This extra effort is precisely the real culprit that quietly steals the most time in writing.

For this, I want to use Claude Code's Agent Skills. Skills is a mechanism for giving the AI a writing manual tailored specifically to you. It is like handing a comprehensive operations manual to a new hire just once, rather than explaining everything verbally every time.

Benefit 1: The writing style stays consistent every time

If you write rules for style, structure, and terminology in the Skill, the result will have the same quality no matter who asks or when. The team's articles gain consistency, and the problem of “the tone differs depending on who is in charge” disappears.

Benefit 2: Instructions become dramatically shorter

With just one instruction—“Make it a Qiita article”—all the lengthy prerequisites are conveyed. Every prompt changes from a long, incantation-like passage of text into a one-phrase cue.

Benefit 3: Reuse it and grow it over time

Because a Skill is just a file, it is easy to improve and share. The more successful writing patterns you add, the more the AI grows like your alter ego.

Now, take the first step

Even without any programming knowledge, Skills can be created simply by “writing the rules in prose.” **Turn writing from a repetitive task into an asset to nurture.** Starting today, let's begin growing your own dedicated writing partner.

There are six instances of bold text. This is a stress test packed with expressions that sound like promotional copy written by AI, including “dramatically,” “a long, incantation-like passage,” “your alter ego,” and “an asset to nurture.”

The prose is still just as dense as ever, isn't it?

Output from yomiyasu v2.5.0

This is the output of v2.5.0.

Have you had enough of writing the same instructions over and over every day?

Every time you write an article, have you asked AI things like “use desu/masu style,” “make it beginner-friendly,” and “use this heading format” all over again? This extra work is what takes up the most time in writing, while going largely unnoticed.

Benefit 1: The writing style stays consistent every time

If you write rules for style, structure, and terminology into the Skill, whenever and by whomever it is requested, the result will be of the same quality. The team's articles will gain consistency, and the problem of "the tone varies depending on the writer" will also be resolved.

Benefit 2: Instructions become shorter

With just the words "Turn this into a Qiita article," all the lengthy prerequisites are conveyed. Every prompt changes from a long, spell-like incantation into a one-phrase signal.

Benefit 3: Reusable and improvable

Since a Skill is simply a file, it is easy to improve and share. The more successful writing approaches you add, the closer the AI's output comes to your writing style.

The First Step

Even without any programming knowledge, all you need to do to create Skills is "write the rules in words."Let's turn writing from a repetitive task into an asset you can grow. Start nurturing your own dedicated writing partner today.

The original heading and the six bolded sections were left unchanged. Meanwhile, "dramatically shorter" was changed to "instructions become shorter," and "your alter ego" was made more specific as "your writing style."

I made some adjustments to the metaphors, but left quite a lot of decorative styling in place. That was the output from v2.5.0.

Output from yomiyasu v3.0.0

Here is the actual output.

Wouldn't you like to eliminate the effort of writing the same instructions every time?

Every time you write an article, do you find yourself asking the AI again to "use the desu/masu style," "write for beginners," and "use this heading format"? This tedious effort quietly steals the most time from writing.

This is where Claude Code's Agent Skills come in. Skills are a mechanism for giving the AI your own "writing manual." Instead of explaining everything verbally to new team members every time, you hand over a comprehensive work manual just once.

Benefit 1: Consistent writing style every time

If you write the rules for style, structure, and terminology into the Skill, whenever and by whomever it is requested, the result will be of the same quality. The team's articles will gain consistency, and the problem of "the tone varies depending on the writer" will disappear as well.

Benefit 2: Instructions become dramatically shorter

Benefit 3: Reusable and improvable

Since a Skill is simply a file, it is easy to improve and share. The more successful writing approaches you add, the more the AI grows like your own alter ego.

The First Step

Even without any programming knowledge, you can create Skills “just by writing the rules as text.” You can transform writing from a repetitive task into an asset you nurture. Starting today, let’s begin nurturing your very own writing partner.

The initial heading and the reader-facing callout changed, and only Agent Skills remained in bold. Yet metaphors such as “becomes dramatically shorter,” “a long, spell-like passage,” and “your own alter ego” remained.

I made changes to the headings and formatting, but the metaphors surprisingly remained. That was the output of v3.0.0.

When you line up the three generations, you can see that the places that were revised differ.

For the same source text, the previous v1.0.1 and this round’s v2.5.0 and v3.0.0 made the following revisions.

Kept in all 6 locationsKept only Agent Skills

Point of comparison	v1.0.1	v2.5.0	v3.0.0
Initial heading	No change	No change	Changed to “Wouldn’t you like to eliminate the effort of writing it every time?”
“The real culprit stealing your time”	“The reason your time is being stolen”	“The reason the most time is taken from you without your noticing”	“Is stealing your time”
Bold	Removed from all 6 locations
“Becomes dramatically shorter”	“Becomes much shorter”	“The instructions become shorter”	Left unchanged
“A long, spell-like passage”	“A long block of instructions”	Left unchanged	Left unchanged
“Your alter ego”	Remained “your alter ego”	Changed to “your writing style”	Changed to “your own alter ego”
“An asset you nurture”	Changed to the description	Changed to the reader-facing callout	Changed to the description
“Come, take the first step”	No change	Changed to “the first step”	Changed to “the first step”
“Only for You”	No change	No change	Changed to “Only for Yourself”

v1.0.1 removed bold formatting and some forceful metaphors, while retaining headings and direct appeals to the reader.

v2.5.0 kept all bold formatting while revising “dramatic” and “double.” Rather than removing stylistic flourishes, it has shifted toward making ambiguous expressions more concrete.

v3.0.0 revised the opening heading and “Well, let’s take the first step.” It made more changes than v2.5.0 to headings intended to move readers forcefully.

That said, “dramatically shorter,” “a long passage like an incantation,” and “your own double” were left in place. Since the official validation had reported an increase in the amount of change from the original text, I thought it might eliminate AI-like expressions wholesale as well. It wasn’t quite that simple.

Not all bold formatting was removed; only the feature name “Agent Skills” remained bold. It seems decisions were based on what was being emphasized, rather than indiscriminately stripping out stylistic flourishes.

How did Japanese proofreading change with the update?

What these results revealed was that the targets of the revisions changed from version to version, rather than all revisions being applied uniformly more strongly than before.

v3.0.0 kept the original structure almost intact while toning down promotional-sounding headings and direct appeals to the reader. On the other hand, the “dramatic” wording that v2.5.0 had revised was left unchanged. It would be better not to conclude, based on the version number alone, that it will “remove AI-like wording more aggressively than before.”

In this passage, v2.5.0 stood out for making metaphors more concrete, while v3.0.0 stood out for adjusting headings and direct appeals to the reader. However, each is the result of only one run. The results may be reversed in a different run.

If you’re already using v2.5.0, it would be a good idea to run the text you normally work with through both versions, place the original and revised versions side by side, and then decide whether to update. Comparing not only copy like this, but also specifications and internal documents will make it easier to judge how the parts you want to keep will change.

References

• yomiyasu v3.0.0 release announcement (X)
• yomiyasu (GitHub)
• yomiyasu v3.0.0 changelog (GitHub)
• What was measured with yomiyasu v3.0.0 and its limitations (GitHub)
• I created the Skill “yomiyasu (Yomiyasu),” which makes AI-slop Japanese easier to read at the structural level (Zenn)

Related articles

• The new Japanese proofreading Skill “yomiyasu” was going viral, so I compared three Japanese proofreading Skills using Claude … An article comparing three Skills, including yomiyasu v1.0.1, using the same source text and evaluation method as this time
• How much do the results and token usage differ when using different Claude Code models? I created the same thing using Opus + Sonnet versus using Opus for everything and compared them… An article that also runs Claude Code from Codex and compares the two under the same conditions

Announcement

As a final announcement, IDEA is recruiting engineers to work with us.(They have also resumed hiring for people with no experience! 👀)
Please see the recruitment page for details.

We look forward to receiving applications from you!