Polyglot: go multilingual or not, where to start, and whether the current setup is right
Polyglot is a skill for Claude that decides whether, where and how a website should go multilingual, and audits a multilingual setup that already exists. It asks for evidence of demand in each market, orders the markets, checks hreflang with a crawler and gives every language written exit criteria, including how to archive it cleanly.
One of the 14 skills in the pack
Ask it in your own words
- "Should I translate my site?"
- "Is my hreflang right?"
- "can i just run my whole blog through DeepL and publish it in 6 languages"
- "How do I remove a language without breaking anything?"
What happens next
- It measures what exists. Sitemap URLs, indexed URLs and non-brand queries for each current language. If those are not indexed well, fixing that comes before a new language.
- It decides on evidence. A short series of yes-or-no gates ends in a verdict, and candidate markets are ranked. Every score is tagged measured, reported, estimated or assumed.
- It audits hreflang and translations. hreflang_check.py reads annotations in HTML, HTTP headers and sitemaps. overlap_check.py flags declared translations whose text was never translated.
What you get back
- A decision memo: pilot, not now, one strong language first, or one language with several names
- A market ranking, with the evidence tag on every score
- An hreflang audit: findings by priority, literal evidence, the fix at the source
- A rollout sheet per topic: local query, slug, verified labels, reviewer, status
Run on this very site
The hreflang annotations of the English and French pages, in the pages and in the sitemap.
pages 5, clusters 1, errors 2, warnings 0
ERROR H07 [...]: html and sitemap annotations differ - x-default=[...]
INFO H09 [...]: sitemap: no x-default (optional; add if a language selector or default page exists) [2 page(s)]The pages named a default version and the sitemap did not. The sitemap now says the same thing as the pages.
What it will not do
- It never invents search volume for a language: it writes "not measurable here" and the cheapest way to measure it.
- It cannot judge translation quality: the overlap check only reveals text left untranslated, so a named native reviewer is a condition.
- Its crawlers run no JavaScript and cannot know whether an engine honoured an annotation.
What it needs
Claude Code, Claude.ai, Claude Desktop or Cowork. The decision needs no script. The two crawlers need code execution with Python 3.9+ and web access; without them, a manual check covers a few page clusters.
How to install a skill in Claude
Where it hands over
- Link Weaver: hreflang gives a translated page no in-content link: the links inside each language are measured and designed there.
- Odds Maker: To test a query in the language where the results are hardest, before translating the page.
Questions about this skill
Can I just translate everything with DeepL or AI?
As a draft, yes. Published at scale without review or added value, it is what Google's spam policies call scaled content abuse. The skill asks for a native review and a few complete clusters.
Will it ever tell me not to translate?
Yes, and it always states the alternative: one strong language first. When the foreign-language searches are product names typed by the market you already serve, it keeps a list of aliases and adds no translation.
It comes with the 13 others.
One pack, one payment: the 14 skills, their scripts, templates and reference notes.
One payment, no subscription. In your inbox within a minute, as a rule.