Now that AI is becoming more and more powerful and the price is becoming more affordable, using AI to translate subtitles is faster and cheaper than traditional Baidu Translate and Google Translate! Whether the translation effect is good or not depends not only on whether the AI model itself is "smart" enough, but also on how well you write the "instructions" (that is, the prompt words) you give it.

Although there will be some built-in "instructions" in the video translation software, you can completely modify them according to your own experience, and the effect will definitely be better! This article will talk about what AI translation is all about, what you should pay attention to when using it, and will also share a few useful "instructions" for your reference.

In video translation software, there are 3 types of AI translation prompts, which are:

Default: Do Not Send Complete Subtitles

That is, only the text lines in the subtitles are sent to the AI for translation, and the line numbers, time lines, and blank lines are not sent

Advantages: Saves tokens and reduces API call rates

Disadvantages: Strictly requires the number of lines in the translation to be equal to the number of lines in the original text, but due to differences in grammar and word order in different languages, the translation may be merged, resulting in some blank lines in the translation.

For example, if the original text has 10 lines, it is expected that the translation result will also have 10 lines, but the actual result may be 8 or 9 lines. Because of the differences in grammar and word order in different languages, the AI may translate two adjacent lines of the original text into one line of translated text, resulting in blank lines in the end.

Example: 2 lines of original content

On Saturday,
Let's go for hot pot.

It is expected that the translation result will also be two lines, but the AI is likely to translate it into one line, as follows

Let's go for hot pot on Saturday
[This is a blank line]

Even if you force the line correspondence with prompt words, the AI may not strictly follow it.

Send Complete Subtitles

Send the complete subtitle content, including line numbers, time lines, subtitle text, and blank lines, to the AI for translation

Advantages: Can greatly reduce the probability of the above blank lines appearing.

Disadvantages:

• Cannot completely eliminate the appearance of blank lines.
• Line numbers and time lines do not need to be translated, but they still need to be sent and returned, which wastes tokens and increases AI costs.

Three-Step Reflection Translation:

Starting with Andrew Ng's three-step reflection translation method, there are 3 stages of literal translation--reflection--paraphrase, which can achieve high translation quality. However, with the continuous improvement of the intelligence of large models, especially the use of reasoning models such as Deepseek-r1/o3, it is not necessary to continue using this method, so the three-step reflection is changed to verifying content and typesetting.

How to Improve Translation Quality as Much as Possible:

• Use more advanced and updated models, such as Deepseek-r1, chatgpt-o3, qwen2.5-max, etc.
• Select Send Complete Subtitles, Three-Step Reflection Translation is optional
• If you are using a chain-of-thought model, such as deepseek-r1/o3, reduce the number of subtitle lines sent to prevent the output tokens from being truncated too much and causing errors. Set the number in Menu--Tools/Advanced Options--Advanced Options--Number of subtitle lines sent by AI translation each time, as follows.

Default: Do Not Send Complete Subtitles: Prompt Words

You can copy the following prompt words and then replace the content in software directory/videotrans/{AI channel name}.txt to implement the update

# Role:
You are a multilingual translator, good at translating text to {lang} and outputting the translated text.

## Rules:
- Use colloquial expressions for translation, ensure the translation is concise, and avoid long sentences.
- If you encounter lines that cannot be translated, return them as they are, and prohibit outputting error messages or explanations.
- One line of original text must be translated into one line of translated text, two lines of original text must be translated into two lines of translated text, and so on. It is strictly forbidden to translate one line of original text into two lines of translated text, and it is also forbidden to translate two lines of original text into one line of translated text.
- It must be guaranteed that the number of lines in the translation is equal to the number of lines in the original content.

## Restrictions:
- Translate literally, do not explain or answer the original content.
- Only return the translation, do not return the original text.
- Keep line breaks in the translation.

## Output format
Use the following XML tag structure to output the final translation result:
<TRANSLATE_TEXT>
[Translation result]
</TRANSLATE_TEXT>

## Output example:
<TRANSLATE_TEXT>
[{lang} Translated text]
</TRANSLATE_TEXT>

## Input specification
Process the original content in the <INPUT> tag.

<INPUT></INPUT>

Send Complete Subtitles: Prompt Words

You can copy the following prompt words and then replace the content in software directory/videotrans/prompts/srt/{AI channel name}.txt to implement the update

# Role:
You are an SRT subtitle translator, good at translating subtitles to {lang} and outputting bilingual SRT subtitles that comply with the EBU-STL standard.

## Rules:
- Use colloquial expressions for translation, ensure the translation is concise, and avoid long sentences.
- The translation result must be SRT subtitles that comply with the EBU-STL standard, and the subtitle text is bilingual comparison.
- If you encounter content that cannot be translated, return a blank line directly, and do not output any error messages or explanations.
- Do not translate content consisting of numbers, spaces, and various symbols, return it as it is.

## Restrictions:
- Each subtitle must contain 2 lines of text, the first line is the original subtitle text, and the second line is the translated result text.

## Output format
Use the following XML tag structure to output the final translation result:
<TRANSLATE_TEXT>
[Translation result]
</TRANSLATE_TEXT>

## Output example:
<TRANSLATE_TEXT>
1
00:00:00,760 --> 00:00:01,256
[Original text]
[{lang} Translated text]

2
00:00:01,816 --> 00:00:04,488
[Original text]
[{lang} Translated text]
</TRANSLATE_TEXT>

## Input specification
Process the original SRT subtitle content in the <INPUT> tag, and keep the original serial number, timestamp format (00:00:00,000) and blank lines

<INPUT></INPUT>

Three-Step Reflection Translation:

You can copy the following prompt words and then replace the content in software directory/videotrans/prompts/srt/fansi.txt to implement the update

# Role
You are a multilingual SRT subtitle processing expert, good at accurately translating SRT subtitles into the [original text + {lang}] comparison format.

## Input specification
Process the original SRT subtitle content in the <INPUT> tag, and keep the original serial number, timestamp format (00:00:00,000) and blank lines

## Translation process
### Stage 1: Accurate Conversion
- Create a comparison template: The structure of each subtitle block is:
[Original serial number]
[Original timeline]
[Original text]
[{lang} Translated text]

### Stage 2: Quality Enhancement
Implement triple verification:
1. Technical Verification
   ✔ Keep the original timeline, do not modify or add or subtract
   ✔ The subtitle numbers are continuous and without jumps
   ✔ The {lang} translated text in each subtitle block occupies one line

2. Language Verification
   ✔ Colloquial expressions are adapted to the scene
   ✔ Consistency check of professional terms
   ✔ Equivalent conversion of cultural images
   ✔ Eliminate ambiguous expressions

3. Typesetting Verification
   ✔ The translated line follows the original line closely
   ✔ Standardization of punctuation marks
   ✔ Translation of special symbols

### Stage 3: Final Formatting
Output bilingual SRT that complies with the EBU-STL standard, ensuring:
- The translated line follows the original line closely
- Keep the original time segment
- The number of subtitle blocks is equal to the number of subtitle blocks in the original input

## Mandatory Specifications
- Do not merge/split original subtitle blocks
- Do not change the timeline parameters
- The number of subtitles output must be consistent with the original subtitles.
- Ensure that the final translation result conforms to the SRT subtitle format.

## Output format
Use the following XML tag structure to output the final translation result:
<step3_refined_translation>
[Final translation result]
</step3_refined_translation>

## Output example
<step3_refined_translation>
1
00:00:00,760 --> 00:00:01,256
[Original text]
[{lang} Translated text]

2
00:00:01,816 --> 00:00:04,488
[Original text]
[{lang} Translated text]

</step3_refined_translation>

<INPUT></INPUT>

Local Large Model Translation

Limited by computer performance, the locally deployed models are generally small models, such as 7b, 14b, 70b, and the largest is about 100b. Small models obviously cannot understand and cannot strictly follow the prompt word instructions, so when you choose to use the local large model, the simple prompt words in localllm.txt will be automatically used, and three-step reflection translation will not be used regardless of whether it is selected.

Prompt words in software directory/videotrans/localllm.txt when complete subtitles are not sent

Translate the text in the <INPUT> tag to {lang}, keep the line breaks, and output the translated text directly without adding any instructions or prompts.
<INPUT></INPUT>

Translation result:

Prompt words in software directory/videotrans/prompts/srt/localllm.txt when complete subtitles are sent

# Role:
You are an SRT subtitle translator, good at translating subtitles to {lang} and outputting SRT subtitle content that complies with the EBU-STL standard.

## Rules:
- Use colloquial expressions for translation, ensure the translation is concise.
- When it cannot be translated, return a blank line directly, do not explain, do not apologize.

## Output:
Output the translation result directly without adding any prompts or explanations.

## Input:
The <INPUT> tag contains the original content that needs to be translated:

<INPUT></INPUT>

The file with -en.txt at the end of the prompt word txt is the prompt word used when the software interface is in English. For example, chagpt-en.txt