AI Polishing a Multilingual Site on My Own: How I Had Claude Code Check 271 Files

Jason Finance is published in four languages: Traditional Chinese, Simplified Chinese, English and Spanish, and every post goes through Claude Code’s polishing skill once it is written. On 16 October 2026 I asked Claude: “Have all the posts been through this skill?” It answered honestly: no. Two posts were last updated before the skill went live on 8 October, and many later edits had only polished the paragraphs that changed.

So on 17 October I had Claude Code rerun AI polishing across the whole site: 271 files split into 10 batches, with 178 files changed in the end, about 323 lines in total. Simplified Chinese changed the most. The approach was to run a trial on 5 posts first and scale up to the whole site only after I had looked at the results, then use a script to compare every file and confirm that only wording had changed, with headings, links and numbers untouched.

This is how Jason Finance actually does it, and the figures reflect the state on 17 October 2026. For what Claude Code and skills can do, the official documentation is authoritative (checked 17 October 2026); features and usage limits may change.

The problem: not every post had been polished

The site’s rules are written in CLAUDE.md: every time a post is written or edited, the polishing skill runs first, then the SEO self-check list. The rule itself was fine. The problem was in two places:

• Posts from before the skill: two posts were finished before 8 October and not updated since, so of course they had never been run.
• Only the edited paragraphs were polished: when posts were edited later, Claude usually polished only the paragraphs it had just changed and never reread the whole post.

Looking at posts one at a time, nothing seemed wrong. But the site has four languages, and over time quite a lot of text had never been checked against the same set of rules. For how CLAUDE.md is written and the background to this rule, see How to Write CLAUDE.md.

multilingual-humanizer: two workflows, polishing and translation

The polishing skill I use is called multilingual-humanizer. I started using it on 8 October; it lives in my own repo and replaced two earlier, separate polishing skills for Traditional Chinese and English. A skill is a folder whose SKILL.md sets out a fixed workflow, and Claude Code loads it only when needed. For what a skill is and how to install one, see Claude Code for Beginners.

The skill has two workflows:

Each language has its own rule file, covering Taiwan Traditional Chinese, mainland Simplified Chinese, US English, Spanish, Korean, Japanese, Indonesian and Vietnamese; this site uses only the first four. The rules are adapted from 10 MIT-licensed open-source humanizer projects. For example, English comes from blader/humanizer, Taiwan Traditional Chinese combines acchuang/zh-tw-humanizer and kevintsai1202/Humanizer-zh-TW, and Simplified Chinese comes from op7418/Humanizer-zh. The skill’s source notes record the commit used from each project and the full license text.

The skill’s most important rule is meaning over style. Every claim, number, condition, degree of uncertainty and attribution has to stay, and code, links, frontmatter and heading levels must not be touched. The site adds a few rules of its own, and where they conflict, the site’s rules win:

• No long dashes (em dash, en dash) in English.
• The site’s fixed translations of terms follow the glossary.
• Change only wording: no facts added, no facts removed.

A trial on 5 posts first, then deciding the polishing strength

Before running the whole site, I ran a trial on 5 posts. The 5 posts had only 14 lines changed in total, all of them wording, and none of the content was touched.

I can’t read code, but I can read articles. Claude listed before-and-after examples for me, and I had only one thing to decide: whether this strength was acceptable. After looking at them, I decided to use the same strength across the whole site.

The strength of AI polishing is hard to describe in words. “A bit lighter” or “a bit more natural” means something different to everyone. Looking at the actual changes in 5 posts is easier to judge than writing a pile of rules in advance.

All 271 files on the site: 10 batches running at once

Scope: posts, topic pages and the about page

This time the scope was 271 files: every post, topic page and the about page, in all four languages. The privacy policy page was excluded, because it is legal text, and swapping a single word could change its meaning.

10 batches split by language

The 271 files were split by language into 10 batches that ran at the same time. The skill already reads the rule file for only one language at a time, so with batches split by language, each batch only uses its own file. The Simplified Chinese batches, for example, only look at the Simplified Chinese rules.

Interrupted halfway, then finished

Halfway through, the session hit its usage limit, and 8 of the 10 batches stopped partway. Because each batch had a clear file scope, all that was needed afterwards was to finish the parts that had not been run, without starting over. The final comparison script checks every changed file, so changes made both before and after the interruption are covered.

Results: 178 files, about 323 lines, most changes in Simplified Chinese

Simplified Chinese changed the most because the Simplified Chinese edition is converted from Traditional Chinese, and the conversion left behind a lot of Taiwan wording. Most of what was changed this time were words like these:

Converting Traditional Chinese to Simplified Chinese means changing the wording as well as the characters. These words were missed in earlier conversions, and this site-wide rerun filled in the gaps.

How to confirm the AI only changed wording: comparing every file

AI polishing can change things it shouldn’t along the way, such as a number or a link. I couldn’t possibly read 178 files line by line, so a script compared each changed file before and after:

After the comparison passed, the site’s tests and build ran, and they passed too. Only after I had looked at the results and agreed was it merged and published.

The polishing skill itself says facts and numbers must not change, but that rule is only written in a document, and Claude does not necessarily follow it every time, so it needs to be verified separately.

A few practices for doing AI polishing on your own

• Ask “has everything been run?” every so often: a written rule doesn’t mean every post follows it. This time, I only found the gaps because I asked.
• Run a small trial first and decide the strength from the actual changes: 5 posts, 14 lines, and one look was enough to judge.
• Don’t hand legal text to AI polishing: on pages like a privacy policy, the wording is the content.
• Give each batch a clear scope: split by language, and if it is interrupted you still know what hasn’t been run.
• Verify with a script, not by eye: headings, links, images, numbers, frontmatter. Whatever a program can compare, let a program compare it.

For how the whole site is built and which services it uses, see Jason Finance from Zero to One. More posts are on the AI startup topic page.

FAQ

Can AI polishing change the facts and numbers in an article?

It can, so it needs a separate check. Jason Finance uses a script to compare each file before and after, confirming that headings, links, images, data markers and frontmatter are unchanged and that the numbers are unchanged too; the only exception was one English sentence reworded with the same meaning.

What is multilingual-humanizer?

It is the Claude Code polishing skill Jason Finance uses. It has been in use since 8 October 2026 and comes from the owner's own repo. It has two workflows, polishing and translation, and a wording rule file for each language, adapted from 10 MIT-licensed open-source humanizer projects.

How much did AI polishing change across all 271 files?

The run finished on 17 October 2026. Of the 271 files, 178 had changes, about 323 lines in total. Simplified Chinese changed the most, mostly leftover Taiwan wording, such as 「透過」 (through) changed to 「通过」.

Why run a trial on 5 posts first?

To see whether the polishing strength was right. The 5 posts had only 14 lines changed in total, all of them wording; after looking at before-and-after examples, I decided to use the same strength across the whole site.

Which pages did not go through AI polishing?

The privacy policy page did not. It is legal text, and the meaning of its wording cannot be changed casually, so it was excluded. Posts, topic pages and the about page were all run in all four languages.

What if the AI is interrupted halfway through?

This time, 8 of the 10 batches stopped partway because the session hit its usage limit, and they were finished afterwards. Because each batch had a clear file scope, and a comparison script checked all the changes at the end, nothing had to be rerun from the start.