Customizing AI Translation with Custom Prompts

Smartcat's AI translation can be customized through two mechanisms: AI engine presets and AI Actions. Both are configured in Workspace settings → AI engine presets. Custom prompts allow you to control tone, register, terminology, domain, output format, and quality through natural language instructions to the AI model.

This article explains how to configure these customization options to produce higher-quality, more consistent translations tailored to specific content types, domains, and audiences.

When to use it

Custom prompts are valuable when:

• Your default AI translation output uses the wrong register or tone for your brand

• You need domain-specific terminology or style rules applied consistently

• You want to enforce glossary terms and forbidden terms at the prompt level

• You need different AI behaviors for different content types (marketing copy vs. legal documents vs. product descriptions)

• You want to run multi-step translation with self-review (chain-of-thought) for high-stakes content

How to use it

Workflow 1: Create an AI engine preset

AI engine presets automatically translate all documents in a project using your custom configuration.

To create a preset:

1. Go to Workspace settings → AI engine presets

2. Click Add Preset

3. Select your AI model (GPT-4o, Claude, Gemini, or other LLM-based engine)

4. Write your system prompt (defines the AI's role and behavior)

5. Write your user prompt (contains the text to translate with context)

6. Click Save

To assign the preset to a project:

1. Open your project

2. Go to the Linguistic assets tab

3. In the AI Translation section, select your custom preset for the relevant language pair

Model selection guidance:

More capable models (Claude Sonnet/Opus, GPT-4o, Gemini Pro) follow complex prompts more reliably. Balance capability with volume and cost considerations. For high-stakes content with detailed instructions, use more capable models. For high-volume standard content, simpler models may suffice.

System prompt vs. user prompt:

• System prompt: Defines the AI's role, behavior, and constraints. This is where you set tone, style rules, and conditional instructions.

• User prompt: Contains the actual text to translate, typically using the {SOURCE_TEXT} placeholder.

Workflow 2: Write an effective prompt

Custom prompts use placeholder variables that Smartcat replaces with actual values at translation time.

Placeholder reference:

Example 1: Generic translation preset

A minimal prompt for standard translation:

System prompt:

You are a professional translator from {SOURCE_LANGUAGE} to {TARGET_LANGUAGE}. When given a text in {SOURCE_LANGUAGE}, translate it into {TARGET_LANGUAGE}. Your entire response must be a translation, without any additions.

{IF GLOSSARY_TERMS}Use all translations from the following glossary: {GLOSSARY_TERMS}{/IF GLOSSARY_TERMS}

User prompt:

{SOURCE_TEXT}

Example 2: Domain-specialized preset (pharmaceutical)

For regulated content requiring specific terminology and style:

System prompt:

You are a professional medical translator from {SOURCE_LANGUAGE} to {TARGET_LANGUAGE}. Follow these rules:

• Use formal, clinical language appropriate for healthcare professionals

• Preserve all drug names, dosages, and medical abbreviations exactly as written

• Translate anatomical terms using standard medical nomenclature

• Do not simplify or paraphrase clinical instructions

{IF GLOSSARY_TERMS}Apply these approved terminology translations: {GLOSSARY_TERMS}{/IF GLOSSARY_TERMS}

{IF FORBIDDEN_TERMS}Do not use these terms—replace with approved alternatives: {FORBIDDEN_TERMS}{/IF FORBIDDEN_TERMS}

User prompt:

Translate this pharmaceutical content:

{SOURCE_TEXT}

Example 3: Chain-of-thought preset for high-stakes content

For content requiring self-review before final output:

System prompt:

You are a professional translator from {SOURCE_LANGUAGE} to {TARGET_LANGUAGE}. For this high-stakes content, follow these steps:

Step 1: Translate the source text

Step 2: Review your translation for accuracy, tone, and terminology

Step 3: Identify any potential issues or ambiguities

Step 4: Revise the translation to address any issues

Step 5: Provide only the final revised translation

{FINAL_OUTPUT}

Your final translation goes here. Only the text between these markers will be used.

{/FINAL_OUTPUT}

{IF GLOSSARY_TERMS}Required terminology: {GLOSSARY_TERMS}{/IF GLOSSARY_TERMS}

User prompt:

{SOURCE_TEXT}

Important: When using {FINAL_OUTPUT} markers, both the opening and closing tags must be present. The AI's response between these markers is extracted as the final translation. Missing markers will cause extraction to fail.

Workflow 3: Configure AI Actions

AI Actions run on selected segments inside the CAT Editor, allowing you to transform existing translations.

To access AI Actions in the Editor:

1. Open a document in the Smartcat Editor

2. Select one or more segments

3. Click the lightning bolt icon in the toolbar

4. Choose an action from the menu

Built-in AI Actions:

• Rephrase: Reword the translation while preserving meaning

• Shorten: Make the translation more concise

• Fix grammar: Correct grammatical and spelling errors

• Fix punctuation: Apply correct typographic punctuation for the target language

• Translate with GPT: Re-translate using GPT with glossary enforcement

To create custom AI Actions:

1. Go to Workspace settings → AI engine presets

2. Click Add Preset

3. Select your AI model

4. Give the action a descriptive name (e.g., "Make more formal" or "Adjust for female speaker")

5. Write your system and user prompts

6. Enable Share to assignee if linguists should see this action in the Editor

The {TARGET_TEXT} placeholder:

AI Actions use {TARGET_TEXT} to reference the existing translation. This placeholder is only available in AI Actions, not in engine presets.

Example custom AI Action — "Make more formal":

System prompt:

Rewrite the given {TARGET_LANGUAGE} text to use formal register. Address the reader using formal "you" where applicable. Avoid contractions and colloquialisms.

User prompt:

{TARGET_TEXT}

How fuzzy matches work with custom prompts

When you include the {IF FUZZY_MATCHES}...{/IF FUZZY_MATCHES} block in your system prompt, Smartcat retrieves Translation Memory fuzzy matches at runtime and injects them into the prompt as context.

This is a RAG (Retrieval-Augmented Generation) pattern: the AI receives relevant past translations as reference material, helping it maintain consistency with your existing translation memory while still producing new translations.

The fuzzy match content appears only when matches exist. If no fuzzy matches are found for a segment, the conditional block is omitted entirely.

Requirements and limitations

• LLM-based engine required: Custom prompts require an LLM-based engine (GPT-4o, Claude, Gemini). They are not available with traditional machine translation engines like Google NMT or DeepL.

• Workspace-level configuration: AI engine presets are configured at the workspace level and can be shared across projects.

• {TARGET_TEXT} scope: The {TARGET_TEXT} placeholder is only available in AI Actions, not in engine presets.

• Chain-of-thought markers: When using {FINAL_OUTPUT} markers, both opening and closing tags must be present. Missing markers will cause extraction to fail.

• Prompt complexity: Overly complex prompts can reduce translation quality. Start simple and add complexity incrementally.

• System-level constraints: Custom prompts cannot override system-level translation principles built into Smartcat's AI translation pipeline.

FAQs

Q: What's the difference between an AI engine preset and an AI Action?

A: Engine presets automatically translate all documents in a project. AI Actions are applied manually to selected segments inside the CAT Editor. Both are configured in Workspace settings → AI engine presets.

Q: Can I use the same preset across multiple projects?

A: Yes. Presets are saved at the workspace level and can be assigned to any project via the Linguistic assets tab.

Q: My prompt works for English→Spanish but not for English→Japanese. Why?

A: Prompts behave differently across language pairs due to linguistic and cultural differences. Test each target language individually. See Troubleshooting AI translation quality issues for language-specific guidance.

Q: How do I stop the AI from translating brand names?

A: Use the {IF FORBIDDEN_TERMS}...{/IF FORBIDDEN_TERMS} block in your system prompt, and add the brand names to your forbidden terms list in Workspace settings → Linguistic assets → Glossaries.

Q: Does using a custom prompt cost more Smartwords?

A: No. Custom prompts do not affect Smartwords consumption. Costs are calculated based on source word count and content type, not prompt complexity.

Related articles

• Troubleshooting AI translation quality issues

• AI Translation Overview

• Review and edit AI translations in Smartcat Editor

• Manage customized AI translation engine presets