How to write AI prompts for translation

AI-powered translation is part of many localization workflows, and it will only continue to grow in popularity among those that are looking for fast translations. However, the output is only as good as the input, and the input is your task. With a well-crafted AI prompt, you can improve the translation quality and consistency, especially when you combine it with the right translation management system (TMS).

Below are some things you need to know as you get started with writing AI prompts for translation to get the best possible results from your preferred large language model (LLM).

1. Give it clear, rich context

Context is probably the most important factor in getting high-quality AI translation. If traditional machine translation engines process sentences as isolated strings, LLMs interpret meaning based on surrounding information, audience characteristics, and intent. Consequently, the more context you provide, the more accurate and natural the translation will be.

AI translation quality dramatically increases when it understands who the text is for, why the content exists, where it will be displayed, which constraints must be respected, and how the brand communicates.

What to include in context:

• Target audience. Tell the AI who will read the final text: general consumers, enterprise buyers, medical professionals, internal employees, or others?
• Communication purpose. Is the text meant to inform, persuade, instruct, warn, inspire, or something else?
• Content type. Translations differ by category, so your LLM will need to know whether it translates UI strings, legal agreements, user guides, product listings, or other type of content.
• Industry or domain. There are certain industries like medical, legal, financial, and technical that require much higher precision and specialty vocabulary.
• Locale requirements. AI must be explicitly told which variant of a language (Spanish for Spain vs. Mexico, for example) to produce.

2. Specify output requirements in detail

LLMs can mess things up, so they need explicit formatting instructions to avoid errors. Human translators intuit formatting rules, but AI generally does not. It will freely rewrite, reorder, or “correct” formatting unless directly told not to. Explanations are especially important when you’re dealing with UI strings, code snippets, HTML content, structured data (JSON, XML), placeholders, variables, titles, headings, and lists.

Here’s what you should specify:

• Required format. Tell the LLM if the content must remain in plain text, markdown, HTML, JSON, XML, or other.
• Preservation rules. Here, you should ask the model to keep the placeholders untouched, preserve HTML tags and attributes, avoid adding extra explanations or comments, retain capitalization/punctuation/ spacing, and anything else that may require preservation.
• Structural constraints. Sometimes, you need to maintain sentence count or keep one translation per line, just to give you a few examples. Tell the AI what your constraints are.
• Non-negotiable prohibitions. We kept mentioned that you need to tell the AI what to do, but it’s also important you tell it what not to do. Maybe you don’t want it to modify numbers, or not translate product names. Whatever it is, make sure it knows your don’ts.

3. Define tone, voice, and style with precision

It goes without saying that if you give your AI vague descriptors, it will produce vague results. Many terms are subjective, and to an LLM, they can mean numerous different variations. In real-world localization, tone, voice, and style matter just as much as correct vocabulary. You need to define the tone and style as clearly as possible, otherwise you might see the AI improvising in all the wrong ways.

Tone expresses attitude, such as confident, warm, neutral, playful, technical, conservative, youthful. Voice is the personality behind the message, and brands spend years cultivating it. And style covers structural and linguistic preferences like sentence length, consistency in terminology, pasive or active voice, and so on.

When you’re working with human translators, things are easier because they can sense the nuance instinctively and pick up the mood of the text. AI is not capable of that; it adapts only when told to adapt. Therefore, you should:

• Describe your desired tone in concrete terms.
• Explain your brand’s voice.
• Consider the cultural and regional style variations.
• Provide a textual moodboard through examples.

4. Incorporate termbases, glossaries, and style guides

LLMs don’t inherently know your preferred terminology, so this is yet another thing you need to tell them. Inconsistent terminology is actually one of the most expensive issues in localization. It’s expensive in the sense that it affects product clarity, user trust, and (of course) brand consistency.

Your TMS likely contains at least one of these features:

• Approved termbases.
• Client-specific glossaries.
• Custom dictionaries.
• Style guides.
• Brand voice guidelines.

Your AI prompts should reference (or embed) these. To make sure your LLM uses your preferred terminology, you should provide it with a clar list of approved terms, tell it how it must use the terms, and include any special rules like never pluralize, never translate, and so on. And don’t forget to mark the words that should remain in the source language.

5. Tell the AI how to handle ambiguities

When you’re working with human translators, things are easier because they can ask questions anytime something’s unclear. However, when working with AI, you need to rely on rules. LLMs often encounter words with multiple meanings, acronyms that overlap across industries, or context that hasn’t been fully provided. It’s important you guide it, otherwise they might select an arbitrary interpretation, and you don’t want that.

A few you can prevent ambiguities:

• Choose the meaning most appropriate for the domain. I you don’t mention your domain, the model may pick a general-purpose meaning.
• Ask for clarification instead of guessing. Though LLMs are not usually conversational in automated workflows, you can instruct them to ask questions instead of hallucinating.
• Choose the literal interpretation by default. It’s often safer to go with literal translation if you content allows it. It’s a way of making sure your translation stays close to the source.
• Pick the most common industry usage. You know how for some words, the meaning depends on how the industry uses them. You can instruct the AI to choose the most commonly used meaning in the industry you’re operating in.
• Apply rules for acronyms and abbreviations. Acronyms too have different meanings, so you need to instruct the model to use the most common ones in your industry.
• Log and highlight ambiguous areas for review. Tell the AI that if any part of the text is ambiguous, it should create a short note identifying potential interpretations.
• Use fallback strategies. You can define a multi-step fallback instruction that guides the model through priorities, because this will give the AI a process to follow.

6. Use Few-Shot Examples

Few-shot prompting means giving the model a handful of examples to mimic. You represent your rules through concrete demonstrations, and because LLMs are good at pattern recognition, they can generalize from even two or three examples to produce output that feels coherent.

If your organization already has glossaries, translation memories, a style guide, and previously approved copy, you can embed slices of that content directly into the AI prompt as few-shot examples. But an effective example must be chosen carefully, so here’s what we recommend:

• It must reflect the exact style you want.
• It should be short enough to fit easily into a prompt.
• It must be parallel (source + target).
• The content should be structurally similar to what you expect.
• They should reflect your terminology choices.

If you don’t use few-shot examples, there’s a greater risk the AI might invent stylistic patterns, use synonyms inconsistent with your glossary, ignore formatting rules, and break consistency. On the other hand, if you do use examples, you will likely do fewer post-editing corrections and get a significantly more consistent tone.

Absolute don’ts when writing AI prompts

The biggest mistake is probably being too vague, because the model will likely fill in gaps with its own assumptions. At the same time, you shouldn’t overload it with long and unstructured instructions, as it performs best when the AI prompts are clear and hierarchical.

A common mistake in AI-assisted translation is telling the model to “think like a professional translator” and assuming that this instruction alone will produce high-quality output. This doesn’t actually give the model a usable guidance, and it might choose the wrong wording. To get the output you want, you must supply the context a real translator would request upfront.

You probably don’t know this, but LLMs have what we could call a positional bias: they pay more attention to the beginning of a prompt than the end. That’s why it’s recommended you add the most important instructions in the very first section of your prompt, ideally under a clear header. You can then follow with examples, source text, formatting details, or other preferences.

Wrapping up

Good AI prompting mirrors the workflows that professional translators have relied on for decades. It gives the AI the same inputs a human linguist needs to produce consistent, high-quality work. You don’t have to reinvent the wheel, you just have to follow some simple rules. The main takeaway is this: when prompts are deliberate and well-structured, your LLM can supports linguistic consistency, accelerate turnaround times, and reduce post-editing work.