Typing to ChatGPT and Claude in two languages: fixing the keyboard-layout mess

Why chat made the layout problem worse

• Short messages, sent fast. You type a line and hit Enter. There's no re-reading pass the way there is with an email, so a wrong-layout word ships before you see it.
• Constant switching. A question in your language, a term or a snippet in English, back to your language. In one thread you can change direction ten times.
• Autocorrect doesn't catch it. To the spell-checker, akuo is a misspelled English word, not שלום in the wrong layout. It gets an underline, not a fix.
• It's the same box all day. Chat, coding agent, search — more of your total typing goes through a single text field than ever, so a layout habit that used to bite once a day now bites hourly.

What happens when a wrong-layout prompt is sent

Models are surprisingly good at decoding common wrong-layout words: they've seen a lot of ghbdtn standing in for привет. But it isn't reliable, and it falls apart on names, numbers, code, and anything the model hasn't seen garbled before. In practice one of four things happens: the model guesses right, guesses wrong and answers something adjacent, asks what you meant, or answers in the wrong language. Three of those cost you a turn.

Garbled text also tokenizes worse, so it costs a few more tokens than the word you meant. That's a footnote, though. The real cost is the re-send and the broken train of thought.

Coding agents raise the stakes

With Claude Code, Cursor, Copilot and their peers, a lot of the typing happens in a terminal, and a lot of what you type is an instruction to something that can run commands. A garbled instruction to an agent is a garbled instruction with side effects. Two things help here:

• Terminal awareness. Unfumble knows a terminal from a chat. In a terminal it recovers a command like go typed on the wrong layout; in a chat it leaves the same keystrokes alone, where they're more likely a name.
• Secure input is respected. If you turn on Secure Keyboard Entry in Terminal, macOS blocks every accessibility tool, Unfumble included. That's the intended behaviour.

A practical setup for bilingual AI work

1. Let the switcher follow the language. Unfumble learns which language you tend to write in each app and website, so a chat you usually run in Hebrew and a code host you run in English each get the right default.
2. Leave the address bar alone. Unfumble defaults a browser's address bar to English so links type cleanly, then hands the language back on the page.
3. Fix a whole prompt before sending. If you notice a wrong-layout paragraph before hitting Enter, select it and press Unfumble's shortcut, or right-click and choose Unfumble from the Services menu.
4. Undo when it's a term, not a typo. A product name in Latin letters inside a Hebrew sentence is a legitimate word. Undo the correction once and, if you tell it to stop, it learns that word for that app.
5. Pause when you're deliberately monolingual. Writing a long English document? Pause Unfumble in that app from the menu bar and resume later.

One thing you don't need to change: mixing languages inside a prompt. Models handle mixed-language input well. The layout was the only real problem.

Frequently asked questions

Do ChatGPT and Claude understand text typed in the wrong keyboard layout?

Often, for common words: models have seen plenty of ghbdtn-for-привет in their training data. It isn't reliable, though, and it fails for names, numbers, code and anything unusual. A wrong-layout prompt usually costs you a clarification or a re-send.

Does Unfumble work inside the ChatGPT and Claude apps and websites?

Yes. It works system-wide in standard text fields, including browser chat boxes and desktop chat apps, and in terminals used by coding agents. Secure input fields such as passwords are excluded.

Does Unfumble send my prompts anywhere?

No. Detection runs on your Mac with bundled dictionaries. Nothing you type is transmitted; the only network use is licensing and signed updates. See what the Accessibility permission really allows.

Does it work in Claude Code, Cursor and other terminal-based agents?

Yes. Unfumble is aware of terminals, so a command typed on the wrong layout is recovered while the same keystrokes in a chat are left alone. It can't act in a terminal with Secure Keyboard Entry turned on.