When I started building an AI comment generator, my first instinct was obvious:
Tune the prompt harder.
I tried to make the model sound more casual. Then more specific. Then less robotic. Then more like a real TikTok user. Then less like a marketing intern pretending to be a real TikTok user.
That approach worked a little.
But after a while, it felt wrong.
I was asking AI to invent the shape of a good comment from second-hand assumptions. I was not looking closely enough at the thing I was trying to generate.
The better question was not:
“How do I prompt the model to write better comments?”
The better question was:
“What do real high-engagement comments actually look like?”
That changed the project.
Prompt tuning is second-hand data
Prompt tuning is useful, but it can become a trap.
If I write:
“Make this comment sound natural, casual, and engaging,”
the model has to guess what “natural” means.
If I write:
“Make this comment sound like TikTok,”
the model usually reaches for a vague internet voice. Short sentences. A little humor. Maybe an emoji. Maybe a phrase like “this is so real.”
That can produce something passable, but it is still built on a stereotype of the platform.
It is second-hand data.
The model is not learning from what people actually liked, replied to, argued with, or remembered. It is learning from my opinion about what I think people like.
That felt too weak for a product.
So I started collecting real comments
For this project, I started collecting high-engagement short-form video comments as research material.
The goal is to keep expanding this into a much larger dataset, eventually tens of thousands of high-like and high-reply comments across different niches.
That matters because “good comment” is not one thing.
A good comment on a fitness transformation video is different from a good comment on a gaming clip. A good comment under a storytime video is different from a good comment under a creator advice video.
If I only tune prompts, I miss that.
If I look at real comments, patterns start to show up.
One high-liked comment says more than a generic prompt
One comment in my dataset had 243,581 likes:
“Everyone talking about Meryl, however, no one pointed out Anne's shaky voice after Meryl's glare, that's top notch acting right there.”
That comment works because it is not just praise.
It does three things:
- It notices a specific detail
- It points out that other viewers missed it
- It makes a judgment
That is much stronger than:
“Great acting, such an amazing scene.”
The second comment is positive, but empty. It could be posted under almost any clip.
The first comment proves the viewer actually saw something.
That is the kind of pattern I want the product to learn from.
The useful signal is in the pattern, not the exact sentence
I do not want to copy high-liked comments.
That would be useless and wrong.
The useful part is the structure.
For example, this pattern appears again and again:
“Everyone is talking about X, but no one is mentioning Y.”
That is a powerful comment shape because it creates a small discovery moment. The commenter is not just reacting. They are reframing what everyone else is watching.
Another common pattern is:
“I thought X was the point, but Y is what got me.”
That works because it has a turn.
Another one:
“As someone who has been in this situation, this part is painfully accurate.”
That works because it adds lived context.
These are not magic phrases
They are interaction patterns.
That is what I missed when I was only tuning prompts.
The product lesson
When I started working on Comment Generator Pro[www.commentgenerator.pro], I thought the hard part would be generating enough comment variations.
Now I think the hard part is building the right input system.
A comment generator should not just produce text. It should protect the user from producing generic text.
That means the tool needs to be grounded in real comment behavior.
Not just prompt tricks.
Not just tone sliders.
Not just “make it casual.”
Real data changes what you notice.
High-like comments are often specific, but not polished.
They are emotional, but not always positive.
They are short, but not empty.
They often contain a twist, a missed detail, or a personal admission.
That is hard to invent from a prompt alone.
Top comments (0)