Lately, I have been using LLMs for a lot of development work.
Sometimes I ask something in English. Sometimes in Bangla. And quite often, I end up mixing both.
That made me think about something:
Does an LLM actually perform better when we talk to it in English?
Especially for us as Bangla-speaking developers, this is a pretty practical question.
We are comfortable explaining an idea in Bangla, but most of the technical world around us—documentation, GitHub, frameworks, error messages, API docs—is in English.
So which one should we use?
After using LLMs for coding and building things, my answer is:
The language matters, but the way you structure the prompt matters much more.
Bangla vs English: Is English Really Better?
For technical tasks, I often prefer English.
Not because an LLM cannot understand Bangla, but because most programming concepts are already expressed in English.
For example, this prompt is perfectly understandable:
"একটা responsive navbar বানাও যেখানে mobile এ hamburger menu থাকবে।"
But this one gives slightly more technical context:
"Build a responsive navbar with a mobile hamburger menu. Use React and Tailwind CSS, and make sure the navigation is keyboard accessible."
The second prompt is not necessarily better just because it is English.
It is better because the requirement is clearer.
And that is an important distinction.
A well-written Bangla prompt can easily be better than a badly written English prompt.
For example:
"আমি একটা developer portfolio website বানাচ্ছি। Designটা clean এবং minimal হবে। Hero section-এ name, role আর short introduction থাকবে। এরপর projects, skills এবং experience দেখাতে চাই। Mobile আর desktop দুই জায়গাতেই responsive হতে হবে।"
That's already a good prompt.
There is enough context for the model to understand what we are trying to build.
So I don't think the real question is:
"Bangla or English?"
The better question is:
"Have I explained what I actually want?"
The Problem With Most Prompts
I think one of the biggest mistakes we make with LLMs is giving it too little information.
For example:
"Build me a portfolio website."
Technically, this is a valid prompt.
But what does "portfolio website" mean here?
Should it be dark or light?
Should it feel corporate or creative?
Should there be animations?
What sections do we need?
What stack are we using?
Who is the website for?
The model has to guess all of that.
And when the model starts guessing too much, the result becomes generic.
Now compare that with:
"I am building a personal portfolio website for a frontend developer. I want it to feel clean, modern, and professional—not overly flashy. Use Next.js, TypeScript, and Tailwind CSS. Include a hero section, selected projects, skills, experience, and contact section. Keep it fully responsive."
This is already much better.
Not because it is much longer.
Because it removes guessing.
A Good Prompt Is Basically a Clear Brief
This is how I have started thinking about prompts.
Instead of treating a prompt like a question, I treat it like a small project brief.
Before asking the LLM to do something, I try to give it four things:
Context — What am I working on?
Goal — What exactly do I want?
Constraints — What should it use, avoid, or preserve?
Output — What kind of result do I expect?
For example:
I am building a portfolio using Next.js and Tailwind CSS.
I want to redesign the projects section.
Keep the current color palette and typography. Do not add another UI library.
First explain the layout idea, then provide the component code.
That's a much more useful instruction than:
"Make my projects section better."
One Thing I Have Learned: Don't Try to Write the “Perfect Prompt”
This is probably the biggest mindset shift.
I don't think prompt engineering is about finding some magical sentence that gives you the perfect result on the first try.
Real development usually doesn't work like that either.
You write something.
You see the result.
You find a problem.
You improve it.
LLM interaction is very similar.
For example:
Build a hero section for my portfolio.
Then after seeing the result:
The structure is good, but it feels too generic. Keep the layout, make the headline stronger, reduce the empty space, and add a subtle animation. Don't change the color palette.
Then:
Keep everything the same, but improve the mobile layout and make the CTA buttons more accessible.
That back-and-forth is where a lot of the real value comes from.
Prompting is not only about asking. It's also about giving good feedback.
So, Should We Use Bangla?
Absolutely.
Especially when you are trying to explain an idea that you naturally think about in Bangla.
I personally wouldn't force everything into English just because the tool is an LLM.
A mixed prompt can actually feel very natural:
আমি একটা developer portfolio বানাচ্ছি। Designটা minimal হবে, কিন্তু একটু premium feel থাকতে হবে. Use Next.js and Tailwind CSS. Hero sectionটা visually strong হবে, কিন্তু unnecessary animation চাই না. Mobile responsiveness must be good.
This is not "bad prompting."
It's simply how many of us naturally communicate technical ideas.
The important thing is that the model understands the requirement.
What I Would Recommend
For Bangla-speaking developers, my practical approach would be:
Use Bangla when you are explaining your idea naturally.
Use English technical terms when they are clearer or more standard.
Use English-heavy prompts for complex coding instructions when you feel the technical requirement is easier to express that way.
But don't obsess over the language.
Focus on these instead:
Be specific.
Give context.
Set boundaries.
Explain the expected result.
That's usually going to improve the output much more than simply switching from Bangla to English.
Coming Next
This was just the first part.
In Part 2, I want to go deeper into the actual prompt structure—how I break down a weak prompt, turn it into a useful one, and use that approach for coding, debugging, UI design, and everyday development work.
Then in Part 3, I'll take a real example and show how I would use prompting to build a website step by step instead of asking the LLM to magically generate the whole thing in one shot.