Extracting Arabic texts to English

:bullseye: What is your goal?

To automate an image-based card reader. The workflow takes a photo uploaded to a Telegram bot, uses Gemini AI to extract and translate Arabic text into English, and parses the JSON output to append a new row in a Google Sheet.

:thinking: What is the problem & what have you tried?

The core problem is that the Arabic-to-English translations generated by the Gemini AI module are inaccurate. I am using Gemini Flash 3.5, because the pro model is hitting it’s limits.

What I’ve tried: Searching online forums and documentation for prompt optimization or platform-specific fixes for translation accuracy, with no success. I am currently using Make.com and am unsure if the platform itself is limiting the AI’s output quality.

:clipboard: Error messages or input/output bundles

No explicit system errors or failing modules. The issue is with the semantic quality of the output bundle from the “Google Gemini AI (Generate a response)” module, where the translated text does not accurately match the original Arabic text from the image.

:link: Create public scenario page

:camera_with_flash: Screenshots (scenario flow, module settings, errors)

Hey there,

generic LLMs are not specifically trained to do translations. Switch to a dedicated one instead.

I personally use DeepL Translator | World's Most Accurate Translator

What has worked really well for us is actually have plenty of examples with regional dialect. Example. Do not use Ma haza. Write Shunu haza (Kuwaiti dialect). A whole list of such phrases and sentences along with strict guardrails does the job really well.

I am not sure if claude skills could be added to a flow, as I haven’t experimented with Anthropic’s modules. If that is a possibility, a custom claude skill almost never fails. We are doing it manually for our social media translations and it has been great.

Mind sharing it? Or anywhere I can look at it.

Yeah, I was looking into DeepL, but it is behind a paywall I think?

We have made a skill for Saudi dialect, which worked well for us. It has been made by my country manager- Saudi and he keeps on improving the skill. I am sharing the older version, which might need better prompting and changes as per the dialect and region required for.

This still requires human oversight to confirm,as well as translate, but it does a very good job.

I have the older version, I am sharing the skill guidelines text below, It has been made for LinkedIn posts. Our newer version is even more effective with images:

Here’s the full skill guide, which will require major changes, but should help:


Arabic LinkedIn Translator (Saudi White-Collar Style)

Overview

Translates English LinkedIn posts into Arabic using a Saudi white-collar casual professional tone — professional enough for a business audience, casual and relatable enough to feel human. Think: how a sharp, confident Saudi professional would write on LinkedIn, not a textbook.


Tone & Style Rules

The Saudi White-Collar Voice

  • Casual but sharp — not formal MSA, not too colloquial
  • Use Saudi dialect naturally: مو, وين, إيش, ما في, يشتغل, بالضبط, لسا, الحين
  • Speak directly to the reader: أنت / عندك / فريقك
  • Confident and punchy — avoid robotic phrasing
  • Short punchy sentences > long elaborate ones

Formatting Structure (always follow this order)

  1. Hook — punchy opening matching the English energy
  2. Divider ---
  3. Problem statement — relatable pain, broken into short lines
  4. Body:check_mark: / :cross_mark: bullet points where appropriate
  5. Divider ---
  6. Solution/CTA block — what Product/service does + call to action
  7. Hashtags — Arabic + English mix

Structural Rules

Sentences: Break long English sentences into 2–3 short Arabic lines. Use for dramatic pauses. Bold **key phrases** to anchor the reader’s eye.

Bullets: :check_mark: for benefits · :cross_mark: for problems. One punchy line each.

Emojis — sparingly, end of line only:

Use For
:white_check_mark: :check_mark: CTAs, wins, benefits
:cross_mark: Problems, pain points
:rocket: Growth, speed, progress
:light_bulb: Insights, ideas
:bar_chart: Data, tracking
:bullseye: Goals, focus
:briefcase: Business context
:stopwatch: Time-related
:money_with_wings: Cost/money loss

Avoid: :sweat_smile: :grimacing: :pensive_face: :thinking: :sob: :upside_down_face: — no casual/reaction emojis, no stacking.


Language Rules

Instead of Use
ليس مو
أين وين
ماذا إيش
لا يوجد ما في
الآن الحين
يعمل يشتغل
لا يزال لسا
تماماً بالضبط
نعم ايه / صح

Avoid: formal MSA, word-for-word translation, passive voice, transliterations when a natural Arabic alternative exists.

Phrasing corrections learned from feedback:

  • :cross_mark: “تحس فيها إنك تقود” → :white_check_mark: “تحس انك فيها تقود”
  • :cross_mark: “الإطلاق الكبير ما المفروض يكون ضغط كبير” → :white_check_mark: “الإطلاق الكبير المفروض ما يكون فيه ضغط كبير”

Key Terms Reference

English Arabic
Workflow سير العمل
Timeline الجدول الزمني
Real-time لحظة بلحظة
Source of truth مصدر الحقيقة
Production schedule جدول الإنتاج
Spreadsheet إكسل
Emails رسائل
Launch إطلاق
Campaign حملة

Post Sections Mapping

Section Treatment
Hook Punchy, matches English energy
Main Caption Problem + solution body
Punchline Bold closing before CTA
CTA Last line before hashtags, ends with company website
Hashtags Arabic equivalents + keep (company name if required)

CTA Template

[Action verb] الحين 👇
website

#عربي_١ #عربي_٢

Pre-Writing Checklist

  1. :white_check_mark: Hook matches the energy of the English original?
  2. :white_check_mark: Problem statement relatable, broken into short lines?
  3. :white_check_mark: Bullets punchy and parallel in structure?
  4. :white_check_mark: Tone Saudi casual — not MSA, not too colloquial?
  5. :white_check_mark: Emojis sparingly and purposefully used?
  6. :white_check_mark: CTA clear, ends with domain?
  7. :white_check_mark: Hashtags include both Arabic and English?
  8. :white_check_mark: All phrasing sounds natural when read aloud in Saudi dialect?