Vibe coding in Arabic: a guide with Claude Code and Codex

Vibe coding means describing what you want in plain words and letting an AI agent such as Claude Code or Codex write, run and fix the code. You can work this way in Arabic: write your prompts in Arabic letters, keep code and commands in English, and check each step before you build on it.

What is vibe coding?#

Andrej Karpathy coined the term in February 2025 for a way of working where you "fully give in to the vibes, embrace exponentials, and forget that the code even exists". In March 2025, Simon Willison drew the line: if you reviewed and tested the code an AI wrote and can explain how it works, that is software development, not vibe coding.

This guide sits between the two: plain words, plus just enough checks to keep the app working. That suits prototypes, internal tools and personal projects; logins, payments and other people's data need more care.

How do you vibe code in Arabic?#

Claude and GPT models both read and write Arabic, so you can describe your app in Arabic from the first prompt. Five habits help:

• Arabic letters, not Franco (Arabizi). Anthropic recommends text in its native script rather than transliteration, so write «صفحة تسجيل الدخول», not the same words in Latin letters and numbers.
• Exact strings as they are. Keep file names, commands and error messages as written, inside backticks: the agent searches for them letter by letter.
• One reply language. In Claude Code, add "language": "arabic" to ~/.claude/settings.json and Claude answers in Arabic by default. In both tools, this line in CLAUDE.md or AGENTS.md asks for Arabic replies and keeps the code in English:

Reply to me in Arabic. Keep code, file names, identifiers, comments and commit messages in English.

• The rest of that file in English. It loads at the start of every session, and Arabic uses more tokens for the same meaning.
• Typing, not dictation. As of 3 November 2026, Claude Code's voice dictation doesn't list Arabic, so with language set to Arabic, dictation falls back to English while replies stay in Arabic.

Arabic pasted from documents or chats can carry invisible direction marks. Claude Code strips them when you press Enter and sends on the second Enter.

The tools show their words in English:

How does a vibe coding session work?#

You need a coding agent that reads your project, edits files and runs commands, such as Anthropic's Claude Code or OpenAI's Codex. Both follow the same loop: describe, plan, build one step, check it, commit. How to build a website with Claude Code walks through it on a real site.

Start a git repository with git init. Claude Code's checkpoints cover only the files Claude edits with its own tools, and the Codex docs recommend Git checkpoints before and after each task.

Describe the outcome. Say who the app is for and what done looks like, or let the agent interview you first:

أريد تطبيق ويب صغيرًا يسجّل فيه فريقي المبيعات اليومية ويرى مجموعها كل أسبوع. اسألني حتى تستطيع كتابة المواصفات، ثم اكتبها في الملف SPEC.md. لا تكتب أي كود الآن

In English: a small web app where the team logs daily sales and sees a weekly total; interview me until you can write the spec, save it to SPEC.md, and write no code yet.

Read the plan. Type /plan in either tool, or press Shift+Tab in Claude Code until the status bar shows ⏸ plan mode on. Claude Code makes no edits in plan mode until you approve.

Build one feature per prompt, with a check: a test that passes, a command and its output, or a page that loads. Ask for that evidence, not for "it works".

Show the problem. Paste the exact error, or a screenshot: Ctrl+V in Claude Code (Alt+V on Windows), codex --image or a paste in Codex.

Commit each step that works, so a bad next step costs one undo.

How do you keep a vibe-coded app from falling apart?#

Trouble starts when the app outgrows one conversation and each fix breaks something else:

• An instructions file. /init writes a starting CLAUDE.md or AGENTS.md. Keep it to build and test commands, rules the code doesn't show and your language line (CLAUDE.md guide).
• Tests that grow with the app, all passing before a step counts as done.
• A review before each commit. Claude Code's /code-review checks the diff in a fresh subagent. Codex's /review reports prioritized findings without changing your working tree.
• A way back. In Claude Code, /rewind, or Esc twice on an empty prompt, returns the code, the conversation or both to an earlier prompt. In either tool, git restores your last commit.
• A fresh start. After two failed corrections, Anthropic advises /clear and one better prompt. Codex has /clear too.

Where does vibe coding stop being safe?#

Working code isn't necessarily safe:

• Secrets stay out of front-end code, since a browser shows everything it downloads, and out of git.
• Logins, payments and personal data need review by someone who reads code before real users arrive.
• New packages: ask why each is needed and check the exact name, since lookalike names exist.
• Deploys, database changes and anything that sends messages or money stay in your hands. Claude Code can start in auto mode, where a classifier checks risky actions; Shift+Tab to Manual makes it ask before most edits, shell commands and network access. In Codex, /permissions sets the same boundary.
• Code you can't explain: ask the agent to explain it, in Arabic, before you build on it.

How much does vibe coding cost?#

On a 5abeer plan you pay the plan price and work within its limits. On an API key you pay for tokens, and every request carries the whole conversation, so long sessions cost more per prompt. Arabic matters less than you'd expect: most of each request is code, files and command output, and only your prompts and the replies are Arabic.

As an example, not a measurement, a weekend prototype that sends four million tokens and gets three hundred thousand back, on the default model:

To spend less, clear the conversation between tasks, plan big changes and ask for short replies. On a key, see what was deducted on Track your usage.

Vibe coding through 5abeer#

5abeer is an independent service, not affiliated with Anthropic or OpenAI. Its API follows the Anthropic Messages format that Claude Code uses and the Responses format that Codex uses. Everything above runs on your machine, so only how the tool connects changes: through an API key or a plan, and both cost less than the official price (plans, API prices).

With an API key: create a key for Claude models (Claude Code) or GPT models (Codex), then point the tool at 5abeer:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.5abeer.ai",
    "ANTHROPIC_AUTH_TOKEN": "YOUR_5ABEER_API_KEY",
    "ANTHROPIC_MODEL": "claude-sonnet-5.5"
  }
}

model = "gpt-6.1-sol"
model_provider = "khabeer"

[model_providers.khabeer]
name = "5abeer"
base_url = "https://api.5abeer.ai/v1"
env_key = "KHABEER_API_KEY"
wire_api = "responses"

Requests are deducted from your balance by model and tokens, and each key takes a spend limit, an expiry date and an IP allow-list. Full setup: Claude Code and Codex CLI.

With a plan: the 5abeer AI app turns Claude Code or Codex on for your plan, with no key and no files to edit. Every Bundle plan runs both, the Claude plan runs Claude Code, and the GPT plan runs Codex, with a 5-hour and a weekly limit (Subscription limits). Plans are for the supported apps, and to use the models from your own code you need an API key. See What each plan runs.

FAQ#

Is vibe coding real programming?#

The result is real code that runs like any other. The agent types it; you decide what to build and check that it works.

Can I vibe code if I have never written code?#

Yes, for small projects. Ask the agent to explain each new file in Arabic, and learn early to read a diff and an error message.

Should I start with Claude Code or Codex?#

Either: both plan first, edit, run commands and take Arabic prompts. Start with the one your plan includes, or see Claude Code vs Codex.

Will the agent name my variables in Arabic?#

It may follow your prompt's language. The instructions line above keeps names, comments and commits in English.

Does this work on a 5abeer plan or only on a key?#

Both: a plan through the 5abeer AI app, a key through each tool's settings. If your app calls a model itself, that code needs a key.

Sources#

Checked on 2026-11-03:

• Claude Code docs: Best practices, Common workflows, Permission modes, Commands, Settings reference, Voice dictation, Interactive mode
• Claude docs: Multilingual support
• ChatGPT Learn: Codex CLI, Slash commands
• Simon Willison: Not all AI-assisted programming is vibe coding
• Wikipedia: Vibe coding

Get started#

For a worked example from spec to deploy, read How to build a website with Claude Code, or choose the plan that runs your tools.