So naturally you start using ChatGPT. Or Claude. Or whatever. You type something in Korean, it doesn't look right, you paste it in and say "make this sound more natural" or "check the spelling" or "rewrite this less formally." And it works. Kind of.
The problem is you have to explain yourself every single time. "Don't make it too formal." "Don't use textbook Korean." "Write it like an actual person would text, not like a language teacher." And the result you get depends entirely on how well you phrased that request. Some days you nail the prompt and get exactly what you need. Other days you rush it and the AI hands you something that sounds like it came out of a Korean 101 course.
So you start saving your good prompts. You find the one that consistently gives you natural-sounding Korean, and you keep it in Notes or your clipboard manager. Now you've got a system. Paste the prompt, paste your text, get the result, copy it, Cmd+Tab back to Slack, paste it there.
It works. But it's a lot of tab switching.
At some point you think — why not just keep one chat open with the perfect system prompt and reuse it? That's the smart move, right?
Except if you've actually done this, you know what happens next. The chat gets long. You use it Monday, Tuesday, Wednesday, and by Thursday something feels off. The responses aren't bad exactly, they're just... not the same. A little more formal than before. A little less natural. You can't quite tell why.
Then you realize the chat is 200 messages deep and the model's context is full. It's not following your original prompt anymore — it's following some blurry average of everything you've said over the last week. So you start a new chat. Set it up fresh. Works great again. For a while. Then it degrades. Again.
That's the cycle. The AI itself is fine. I have no complaints about the output when the prompt is right and the context is clean. What I got fed up with is everything around it. The tab switching. The copy-pasting. The re-explaining. The prompt that works great one day and doesn't the next because the chat's too long. Doing this 30 or 40 times a day, every day.
That's not an AI problem. That's a workflow problem.