Poor results with AI translations

Our existing translations are a mix of human translations and machine translations (approximately 5%). The product manager and others are not happy with many of the newer human translations, so they want to re-translate everything via AI on Crowdin. So, I

  • created a branch
  • worked with the PM to set up a glossary
  • re-tagged all the screenshots for that branch
  • added a description to the project for context
  • unchecked TM suggestions in the prompt editor since existing translations in the translation memory (some of which are bad) might be ‘poisoning’ the AI translations
  • set the model to Opus 5.5
  • tweaked the prompt to make some things explicit
Follow the instructions below:
1. Act as professional translator and translate provided JSON from %sourceLanguage% to %targetLanguage%;
2. For %targetLanguage% language use next ICU plural forms: %pluralForms%. Add missing forms and remove excess forms as needed;
3. Do not use translation memory matches as a reference to provide translations, but do use the project description and screenshots to provide translation context.
4. Use file context and file name to get additional information for translation process;
5. File context can be formatted as markdown or raw text;
6. Always strictly refer to the translation glossary. Always use terms' translations from glossary, even if other more commonly used or recognized translations are available. Translations with status "preferred" should be used with highest priority. Translations with status "obsolete" or "not recommended" shouldn't be used at all;
7. Parse the provided JSON into an array, translate the 'text' property of each object in the array to %targetLanguage% according to the instructions using terms, translation memory matches and file context, do not ignore html markup in the text, if some text cannot be translated then put '' string in 'text' property, serialise the resulting array into JSON, and send that JSON as a response;
8. Do not skip objects in JSON and do not add new objects;
9. Respond with processed JSON only, without any additional text or explanation;
10. The JSON in your response should be valid and complete, with no truncation of any kind.
11. Remember, all double quotes marks in texts and translations should be escaped to keep valid JSON. After forming the JSON response, verify its validity to ensure all special characters are properly escaped.

JSON: 
```
%strings%
``` 

File name: ```%fileName%```

File context:
``` 
%fileContext%
```
Project name: ```%projectName%```

Project description:
``` 
%projectDescription%
```
Style guides:
``` 
%assignedStyleGuides%
```

The translations are still not to the (Italian) PM’s liking.

Any suggestions?

Hello!
There are a few things that should help here:

In an advanced prompt, the instructions in the prompt text don’t send any data on their own. Screenshots and glossary terms are sent to the AI only when the “Screenshots” and “Glossary terms” options are turned on in the prompt settings, in the same place where you unchecked “TM suggestions”. Please make sure both are on. Otherwise steps 3 and 6 point to context the AI never receives.

The biggest lever for meeting your PM’s expectations is a style guide. Your prompt already includes %assignedStyleGuides%, so once a style guide is assigned to Italian (or to all languages), its AI instructions are added to every request. That’s the right place for rules such as formal or informal address (Lei or tu), tone, whether to keep English terms or use Italian equivalents, capitalization and punctuation. Only the AI instructions text is sent, up to 10 KB.

Also, step 7 still mentions “translation memory matches”, which contradicts step 3, so you may want to remove that part. %fileContext% is filled from the description in each file’s settings. If that’s empty, it’s generated from the file content only when the “AI-Generated File Context” option is on, otherwise the AI receives “N/A”. For short or ambiguous strings that aren’t tagged on any screenshot, a string context helps a lot, since it’s sent along with each string.

Thanks, I’m developing a style guide now with the PM and others, and :man_facepalming: that I missed the mention of translation memory in the prompt. Is there a way of seeing everything that’s being sent to the model? I don’t want to see the screenshots, of course, but I would like to see all the text being sent.

Hi there,

At the moment, there’s no way to see the full text sent to the AI model for a specific request. Crowdin doesn’t store the compiled request, and AI Logs only show details such as the model, status, token usage and cost.

What gets sent is controlled by the prompt text and by the context options in the prompt settings, such as TM suggestions, Glossary terms, Style guides and screenshots. That’s also why you didn’t find translation memory mentioned in the prompt text. When the TM suggestions option is enabled, TM matches aren’t added to the prompt as a separate section. They’re attached to each string in the string data sent to the model (the %strings% placeholder), and glossary terms work the same way when the Glossary terms option is enabled. So translation memory is used even though the prompt text doesn’t mention it.

If it helps with your style guide, one of your developers could also get the full prompt template, including the built-in instructions, through the Crowdin API. Please note that this shows the template with placeholders only, not the actual strings, TM matches or glossary terms used in a specific request.

Please let me know if you have any other questions!