Translation prompts: how to create effective prompts for better AI translations

Creating a good translation prompt is one of the most effective ways of improving the results obtained with artificial intelligence tools. However, simply asking the tool to “translate this text” is rarely enough when the goal is a translation that is clear, consistent, technically correct and suited to the target audience.

Digital face with a grid of dots and the title in English “Prompt Translation: How to Create Effective AI Prompts”.

Dokutech’s experience shows that the quality of a translation produced with AI depends on several factors: the context provided, the language pair, the type of text, the terminology, the expected format, the model used and the human review applied at the end of the process.

To assess this reality in practical terms, Dokutech developed and tested a library of translation prompts with different levels of detail, applied to several content areas. The results point to one essential conclusion: AI can speed up and support translation work, but it only produces better results when it is well directed.

What we're going to explore:

What is a translation prompt?

A translation prompt is the instruction given to an AI tool to translate a text from one language into another.

In a simple context, the prompt can be just: Translate this text from English into European Portuguese.

But in a professional context, this instruction is far too vague. The AI does not automatically know the target audience, the intended tone, the approved terminology, the type of document, the language variant or the formatting rules that must be respected.

An effective translation prompt should state, at the very least:

  • the source language;
  • the target language;
  • the intended language variant;
  • the type of text;
  • the target audience;
  • the tone of voice;
  • the preferred terminology;
  • the elements that must not be translated;
  • the expected format;
  • the need to preserve the original meaning without omissions or additions.

The clearer the instruction, the higher the probability that the AI will produce a useful translation at the first attempt.

Four people look up under the question in English “What Is a Translation Prompt?”.

Why does the prompt influence translation quality so much?

Artificial intelligence does not just translate words. It interprets instructions, context and linguistic patterns. That is why two different prompts can generate very different translations from the same text.

A basic prompt can produce a literal, unnatural or inconsistent translation. A more complete prompt, on the other hand, can direct the AI to preserve terminology, adapt the tone to the target audience and maintain the structure of the original document.

In the tests evaluated by Dokutech, more detailed prompts tended to improve results in specialised texts, above all when they included terminology instructions, context of use and quality criteria.

However, the results also showed that more detail does not always mean better performance. In some cases, overly complex prompts generated less consistent results. This reinforces an important idea: a good prompt is not necessarily the longest one, but the clearest, most specific and best suited to the task.

What Dokutech tested in this case study

Dokutech evaluated the use of AI prompts across different translation tasks and language services.

The working base included:

  • 30 defined prompts;
  • 3 levels of detail per prompt;
  • 3 AI tools/models tested;
  • around 270 tests carried out;
  • content spanning technical, medical, marketing, SEO, terminology, post-editing, review, localisation, subtitling and interpreting-preparation tasks.

    Each prompt was tested at different levels:

Level 1: basic prompt.
Simple instructions, focused only on translating the text.

Level 2: intermediate prompt.
Instructions with more context, including type of text, target audience, tone and some linguistic rules.

Level 3: advanced prompt.
Detailed instructions, with quality criteria, execution steps, terminology, style notes and formatting rules.

The results were assessed on the basis of errors such as incorrect terminology, omissions, undue additions, meaning errors, syntax problems and unauthorised changes to the original content.

Main conclusions of Dokutech's tests

1. A more complete prompt tends to improve the translation

The tests showed that prompts with context, purpose and clear instructions help the AI produce more suitable translations.
This was particularly evident in technical, medical and specialised texts, where translation depends on terminological precision and consistency.
When the AI received only a generic instruction, errors were more frequent. When the prompt included context, target audience and terminology, the results improved.

2. Terminology remains the biggest challenge

One of the most recurrent problems in the tests was the incorrect translation of technical terms.
In areas such as automotive, industrial machinery, medical equipment or technical documentation, one badly translated word can change the meaning of the text and compromise the final quality.
That is why a good translation prompt should include glossaries, preferred terms or clear terminology instructions whenever the content is specialised.

3. Glossaries improve consistency

Including glossaries or approved terms proved to be one of the most useful practices for improving translations with AI.
When the prompt indicated how certain terms should be translated, the AI tended to maintain greater consistency throughout the text.
This practice is especially important for companies that use their own terminology, recurring technical documentation or brand-specific language.

4. AI can omit parts of the text

The tests identified situations in which the AI did not translate every segment of the original text, above all where there were repetitions, long sentences or complex structures.
This is a relevant risk in professional translation. A translation can sound fluent and yet be incomplete.
That is why the prompt should include instructions such as:
| Translate all the content of the original text, including repetitions, units, tags, headings and apparently duplicated segments. Do not omit information.

5. AI can add content that was not requested

Another problem observed was the addition of headings, explanations or rewordings that did not exist in the source text.
This behaviour can be useful in copywriting tasks, but it is problematic in translation, above all when fidelity to the original is essential.
An effective prompt should make clear that the AI must not add information, comments or interpretations that are not present in the original text.

6. Human review remains indispensable

The main conclusion of the case study is clear: AI can support, speed up and improve parts of the translation process, but it does not replace human review in demanding professional contexts.
Even where the results were good, there remained risks of incorrect terminology, omissions, changes of meaning or unsuitable stylistic choices.
The best approach is to combine AI with specialised linguistic review. The AI generates a first version or supports post-editing; the human professional validates terminology, meaning, tone, consistency and suitability for the client.

How to create an effective translation prompt

A good translation prompt should be structured like a briefing. The more relevant the information provided, the better the AI’s ability to meet the objective.

1. State the language pair

Always specify the source language and the target language.
Example:
| Translate the following text from German into European Portuguese.
If the language variant matters, state it clearly. “Portuguese” can mean European Portuguese, Brazilian Portuguese or another variant.

2. Define the type of text

The AI needs to know whether it is translating a technical manual, a sales brochure, an SEO article, a medical questionnaire, a subtitle or marketing content.
Example:
| The text is a technical manual for operators of agricultural machinery.
This information helps the AI choose the tone, the terminology and the level of formality.

3. Explain the target audience

The translation must adapt to whoever will read the text.
A manual for specialised technicians requires different language from a brochure for end clients. A medical questionnaire for patients demands clarity and accessibility. A B2B text demands precision and professionalism.
Example:
| The target audience is maintenance technicians with knowledge of the industrial sector.

4. State the tone of voice

The tone can be technical, instructional, commercial, formal, simple, institutional or humanised.
Example:
| Keep a professional, clear and instructional tone.
This instruction avoids translations that are too literal or misaligned with the purpose of the content.

5. Provide preferred terminology

Whenever critical terms exist, include a glossary in the prompt.

Example:

| Use the following terminology:

  • • “control unit” = “unidade de controlo”
  • • “maintenance” = “manutenção”
  • • “trigger” = “disparo”
  • • “patient” = “paciente”

This practice reduces inconsistencies and improves the final quality.

6. State what must not be translated

Many texts contain brands, URLs, codes, placeholders, tags or variables that must stay unchanged.
Example:
| Do not translate URLs, product names, codes, tags in square brackets or elements marked as [NOTRANSLATE].
This instruction is essential in technical translation, software, localisation and content with specific formatting.

7. Ask for the original structure to be kept

If the format matters, the prompt should say so.
Example:
| Keep the structure of the original text, including headings, lists, line breaks, numbering and tags.
This prevents the AI from reorganising the content inappropriately.

8. Forbid omissions and additions

To reduce risk, include explicit instructions.
Example:
| Do not omit information from the original text. Do not add explanations, headings, comments or content that does not exist in the source text.
This rule is one of the most important for professional translation.

9. Ask for a final review

The prompt can end with a quality control instruction.
Example:
| Before finishing, review the translation to make sure it is complete, faithful to the original, terminologically consistent and written in natural European Portuguese.
This step helps the AI check its own output before responding.

Example of an effective translation prompt

Here is a prompt template that can be adapted to different contexts:

| Translate the following text from [source language] into [target language], using [language variant].

| The text is [type of document] and is intended for [target audience].

| Keep a tone [intended tone] and use clear, natural language suited to the context.

| Respect the technical terminology and keep it consistent throughout the text. Wherever applicable, use the following terms:

  • [source term] = [target term]
  • [source term] = [target term]

| Do not translate URLs, product names, codes, placeholders, technical tags or elements marked as untranslatable.

| Keep the structure of the original text, including headings, lists, line breaks and numbering.

| Translate all the content, including repeated segments. Do not omit information and do not add explanations, comments or content that does not exist in the original.

| Before presenting the final translation, review the text to guarantee fidelity to the original, terminological consistency, fluency and grammatical correctness.

| Text to translate:
[insert text]

This template does not guarantee a perfect translation, but it significantly increases the probability of obtaining a first version that is more useful, more consistent and closer to professional requirements.

When should AI be used in professional translation?

AI can support different phases of the linguistic process, above all when it is integrated into a professional machine translation service with specialist review, where the technology is combined with human validation, terminology control and quality criteria.

It can support:

  • creating first versions of a translation;
  • comparing alternatives;
  • identifying inconsistencies;
  • terminology support;
  • post-editing;
  • preliminary review;
  • style adaptation;
  • content localisation;
  • glossary preparation;
  • analysis of tone and clarity.

However, the degree of confidence should vary with the type of content. Technical, medical, legal or highly specialised texts demand more rigorous human review. Marketing texts, localisation or less technical content can benefit from the greater fluency generated by AI, but still need editorial validation.

The limits of AI in translation

Hand with a glowing chip marked “AI”, between the English phrases “Limits of” and “in Translation”.

Dokutech’s tests show that AI has potential, but also clear limitations.

The main risks are:

  • incorrect translation of technical terms;
  • terminological inconsistency;
  • omission of segments;
  • addition of unrequested information;
  • subtle changes of meaning;
  • excessive rewording;
  • lack of alignment with style guides;
  • difficulty interpreting client-specific preferences.

That is why AI should be integrated into a professional translation process, and not used as an automatic solution without quality control.

The question is not just “which tool to use”, but “how to direct, evaluate and review the result”.

What this case study demonstrates

Dokutech’s case study confirms that the quality of translation with AI depends on a combination of factors.

A good prompt improves the results, but does not eliminate the need for linguistic competence. A glossary helps, but does not solve every problem. A model can perform well on one type of text and worse on another. An advanced prompt can be useful for complex texts, but may not be the best option in every case.

Professional translation with AI demands method, testing, comparison and review.

The main lesson is simple: better translations with AI start with better instructions.

FAQs — Learning German with AI

The acronym “FAQ” in orange next to a large white question mark on a grey background.

A translation prompt is the instruction given to an AI tool to translate a text. It should state the language, the language variant, the context, the target audience, the tone, the terminology and the expected format.

To create a good prompt, state the language pair, the type of text, the target audience, the intended tone, the preferred terminology and the formatting rules. You should also ask the AI not to omit or add information.

Not necessarily. Detailed prompts tend to help with specialised texts, but overly complex prompts can generate inconsistent results. The best prompt is clear, specific and suited to the objective.

Not in demanding professional contexts. AI can support the process, but human review remains essential to validate terminology, meaning, style and final quality.

In the tests evaluated by Dokutech, terminology errors were among the most frequent problems, especially in technical, medical and specialised texts.

Conclusion

An effective translation prompt is not just a sentence asking for a translation. It is a structured linguistic briefing, created to direct the AI in a clear and controlled way.

It should state the language pair, the variant, the type of text, the target audience, the tone, the terminology, the format and the essential restrictions. It should also prevent omissions, additions and undue changes to the original content.

Dokutech’s experience shows that AI can be a powerful tool to support translation, review and post-editing. But the best results emerge when the technology is combined with human knowledge, quality criteria and professional validation.

For companies working with technical, medical, commercial or multilingual content, this approach makes it possible to gain efficiency without giving up precision, consistency and trust.

Partilhar

Artigos Relacionados