The more detailed the instructions, the more lost the AI gets; the fix was to ‘provide examples’
Creating content for X, LinkedIn, note, blogs, and podcasts from a single tweet.
I am building an app that does this automatically. Since each medium has different character counts, tones, and reader expectations, simply cutting and pasting the same text doesn’t work. How to get the AI to write this content became the core challenge of this app.
I went through trial and error over 30 times. Maybe even more.
The first trap I fell into was the assumption that ‘the more instructions you add, the higher the accuracy will be’.
The first thing I got stuck on was the method of creating detailed prompts for each posting destination.
In addition to a common prompt, I prepared individual prompts for each platform. I divided the rules for each platform into about five sections, such as ‘for X, use this tone, keep it around 140 characters,’ and so on. ‘For LinkedIn, be more business-like and start with the conclusion.’ ‘For note, start by asking the reader a question and use headings…’
I thought that by stacking up the rules for each medium like this, the accuracy would naturally improve. But when I actually tried it, it wasn’t that simple.
When you prepare both common prompts (rules you want to apply to all media) and individual prompts (rules for each medium), it becomes very difficult to control which one the AI prioritizes.
If you write the common rules too strongly, the individual characteristics disappear. If you write the individual rules too strongly, the common rules stop working. I spent a lot of time in a loop of fine-tuning to try and find a balance.
This is a scene I’ve seen before in my regular work. The more detailed you make the rules and manuals, the more often people on the front lines ask, ‘So, which rule is actually the most important?’ It’s the same with instructions for AI; the more rules you add, the more confused the AI becomes about what it should prioritize. Increasing the amount of information and getting your point across are two different things.
What I stopped doing was ‘writing out instructions,’ and what I started doing was ‘having it read examples’
In the end, I stopped writing detailed formats into the prompts.
Instead, I switched to a method of importing the actual reference text and having it create content based on the atmosphere of that text. In my experience, this has made the generation much more stable.
Instead of ordering it to ‘write like this,’ I provide actual examples and say, ‘I want you to write something like this.’ It is often faster to show someone the finished product than to list a hundred steps.
It was the same for the AI. Rather than a list of detailed rules, high-quality examples help it understand, ‘Oh, I see what you mean.’
Adding 11 rules for phrasing quirks made it sound much more like natural Japanese
Another thing that worked was adjusting the Japanese phrasing.
If left alone, AI-generated text inevitably sounds like a translation. Phrases like ‘can be done,’ ‘in regards to,’ and ‘concerning.’ It produces text that feels somewhat distant, as if English were being mechanically replaced with Japanese.
To fix this, I created 11 paraphrasing rules and added them to the prompt. For example, I specified concrete patterns one by one, such as ‘replace “can be done” with “can do”‘.
As a result, the frequency of translation-style phrasing decreased significantly. This is based on my own subjective feeling that, compared to text generated with the same input on August 7th, the translation-style phrasing listed in the rules had intuitively diminished.
What I learned here is that an AI’s “habits” cannot be fixed by vaguely asking it to “write naturally.” Only when we verbalize what is unnatural and provide a list can the AI begin to avoid those patterns. Just as “please line up your shoes before entering” is more effective than “behave properly,” specific instructions are more effective than abstract expectations.
The core insight that emerged from this trial and error was: “AI writes better when you provide examples rather than giving detailed instructions.”
After more than 30 rounds of trial and error, this is the approach I have settled on for now.
AI writes better when you provide examples rather than giving detailed instructions.
This is a slightly different direction from the mindset that was popular when the term “prompt engineering” was trending. Rather than working hard to construct clever instructions, it is faster to simply show the AI something close to the final result you are looking for.
However, this does not mean that “if you have examples, you don’t need instructions.” Just like when fixing translation-style phrasing, you still need to put into words the specific points you want corrected.
Convey the overall atmosphere and rhythm of the structure through examples, and correct linguistic habits with specific rules. Settling on this three-layer structure, combined with basic prompts, was the conclusion of this trial and error process.
FAQ
Q. Wouldn’t it be better to create both common and individual prompts? I thought so at first and tried to build them that way, but controlling the priority became difficult and it fell apart. It turned out to be more stable to prepare multiple individual examples and keep the common instructions concise.
Q. How many items should I prepare for the phrasing rules? In my case, 11 items reduced the translation-style phrasing from 21 occurrences to just one. The ideal number will vary depending on the content and the type of AI, so it is best to start by listing the phrasing that bothers you.
Q. What kind of text should I choose for the examples? It is best to choose text that you feel is “your own voice.” Rather than using borrowed text, providing something you actually wrote or something you consider ideal will ensure that your own personality remains in the results.
Q. Is this trial and error still ongoing? It is. It took over 30 attempts to settle on the current form, but if I notice new things, if the media increases, or if new AI models are released, I will need to make adjustments again.
#AIUtilization #PromptDesign #ContentCreation #Automation #JapaneseExpression
