Jason JunJason Jun

Using GPT as a language assistant

13 June 2024

Out of all the capabilities of GPT models, my primary use case is still English writing correction. Even after living and working in Australia for over a decade, I still feel uncertain when writing in English. Although I now write mostly in English and rarely in Korean, I still feel the need to have my English writing checked. On the other hand, my Korean writing is getting worse, but I never feel the need to get it corrected. I guess it's the baseline confidence that comes with a native language.

Before ChatGPT, I relied on various online services for help with my English writing. With the exception of Google Translate, all of these services were human-powered. When ChatGPT came out, it replaced all those services for me.

Conversational correction

Initially, I just asked ChatGPT to translate, correct, or improve my writing, just as I would ask a human. However, this process quickly felt repetitive as I had to use the same prompt each time. Also, the answers with this approach are unstable in terms of style. Often, ChatGPT included unnecessary phrases beyond just the correction, such as "Sure, here's the corrected version," which was annoying when I tried to copy the whole response.

Correction by
requestCorrection by request

Although ChatGPT's ability to converse like a human is impressive, my specific use case doesn't necessarily require that interaction style.

UI-based correction

When a process feels repetitive, it's usually a good candidate for automation. Naturally, I quickly built a simple web app using OpenAI's API to solve this. I named it Rephraser, and it was essentially a web UI solution to the repetitive conversation prompts.

The main benefit of this approach was that I could tailor the UI to fit my specific process and access parameters only available via API, such as temperature and frequency_penalty. However, maintaining a web app is a lot of work, and the API fees add to the cost of my existing ChatGPT subscription.

Agent-based correction

In November 2023, OpenAI introduced 'GPTs,' customisable agents for specific use cases. I created one for translation between Korean and English and another for English writing correction. This was convenient because I no longer needed to repeat the same prompt; I just had to choose which agent to interact with. However, switching between different GPTs wasn't always ideal, especially when I needed both in the same writing process. For example, I would translate English to Korean, then write a response in Korean, translate it back to English, rewrite it to match my intended tone, and finally correct any mistakes. This process may require a lot of back-and-forth with copy-pasting.

Then earlier this year, OpenAI added a GPT tagging feature, enabling the use of different GPTs in the same thread. This significantly improved the process because I could use both the translator and corrector in the same thread as needed. Additionally, I could interact with the default ChatGPT to ask about anything during the process. And my favourite part was that the entire process could be done just with the keyboard, without needing to touch the mouse.

Using custom GPTs

First, I created two GPTs with names starting with 'EN...', so they can be easily called by typing @en in the chat.


Both are instructed to only return the result without responding to anything else in the request text.

The translation GPT converts English to Korean and vice versa, making it useful for translating original English text or drafting English from Korean writing.

EN ↔ KR translation
GPTEN ↔ KR translation GPT

The correction GPT corrects grammar or spelling mistakes while keeping the tone and always returns the text in Australian English, which is important to me.

EN correction
GPTEN correction GPT

And there's a handy shortcut for copying the last response to the clipboard: Shift + Command + C on Mac (Ctrl + Shift + C in Windows). This eliminates the need to use the mouse in this correction process.

Copy last
responseCopy last response

Let me demonstrate these in an example scenario.

OS-level writing tools

And this week, Apple finally unveiled their AI features at WWDC 2024, including a new AI-powered writing tool. This tool enables users to rewrite, proofread, and summarise text across both native and third-party apps.

It's exciting to see such rapid evolution in a field that has traditionally been slow and costly. This advancement will greatly benefit those who frequently use a non-native language.