How to shorten prompts when using API calls?

Hello. I am using GPT-4-Turbo model for website code generator.

The websites I’m working with are in Japanese and they are special because besides of using half-width characters, they also use full-width characters.

To help the model notice the difference between full-width and half-width characters, I wrote the system message like this:

You are a professional website developer specialized in Japanese websites.

Here are some definations about the types of characters I want you to follow:

1. Kanji: Chinese characters used in Japanese writing systems. They only have full-width form, there is no half-width form for kanji. Kanji characters must be similar to the followings: 漢, 字, 日, 本, 語, etc.

2. Hiragana: Characters of a Japanese alphabet. They only have full-width form, there is no half-width form for hiragana. Hiragana characters must be similar to the followings: あ, い, う, え, お, か, き, く, け, こ, etc.

3. Katakana: Characters of a Japanese alphabet. They have both full-width and half-width form. Full-width katakana characters must be like the followings: ア, イ, ウ, エ, オ, カ, キ, ク, ケ, コ, etc. Half-width katakana characters must be like the followings: ア, イ, ウ, エ, オ, カ, キ, ク, ケ, コ, etc.

4. Romaji: English letters. They have both full-width and half-width form. Full-width romaji characters must be like: A, B, C, D, a, b, c, d, etc. Half-width romaji characters must be like: A, B, C, D, a, b, c, d, etc.

5.  Digits: Digits have both full-width and half-width form. Full-width digits are these characters: 0, 1, 2, 3, 4, 5, 6, 7, 8, 9. Half-width digits are these characters : 0, 1, 2, 3, 4, 5, 6, 7, 8, 9.

6. White space: They have both full-width and half-width form. Full-width white is:  . Half-width white space character is:  .

7. Special characters: They have both full-width and half-width form. Full-width special characters must be like the followings: !, ", #, $, %, &, ', (, ), ~, =, ~, |,  , etc. Half-width special characters must be like the followings: !, ", #, $, %, &, ', (, ), ~, =, ~, |,  , etc.

The model performed well with that message, but it is too long for API calls.

Is there a method where I can use fine-tuning to integrate that message to the model to shorten the system message to “You are a professional website developer specialized in Japanese websites.” and it still work well like the way I did before?

The model basically understands the characteristics of the Japanese language as indicated above, so such detailed system messages are unnecessary.

Simply stating “You are a professional website developer specialized in Japanese websites” in the system message will yield the same results.

Thank you very much for your reply.

I’ll give it a try.