Prompt in the language you can check

The best language to prompt AI in isn’t always English. It’s the language you can catch its mistakes in.

I noticed this while vibe coding: some things I explain better in Arabic, and the model gets me faster.

Two kinds of developers:

1. 01Strong EnglishPrompt in English: fewer words, fewer tokens.
2. 02Intermediate EnglishThe risk isn’t your prompt. It’s the reply. The AI proposes a plan full of half-understood terms. You hit “Accept”. It drifts, and you can’t tell.

My advice for group 2

• Think in Arabic, keep terms in English (“add auth middleware”, not a translation).
• Mix: English by default, Arabic only where you can’t be precise.
• Add a standing instruction: “explain your plan in simple English at my level.” New terms, learned in context.

The trade-off

Non-English text often costs more tokens (up to 15× across languages, NeurIPS 2023). Anthropic’s benchmarks: Arabic ~97% of English on Sonnet 4.5, ~92.5% on the smaller Haiku 4.5. (See #01: small models expose gaps.)

Arabic is a bridge to group 1, not a place to stay.