Recently, I wanted to turn the articles from my official account into oral broadcast scripts and publish them on other audio and video platforms. After all, since the article is already written, making an audio version would give me a few more distribution channels—the so-called one-fish-multiple-eats, win after win.
The first problem I ran into was that I have absolutely no idea how to write an oral broadcast script. I never studied broadcasting or hosting, and I've never seen what a professional oral broadcast script actually looks like.
But I wasn't planning to learn from scratch either. After all, it's already 2026, and the SOTA models have consumed massive amounts of public information on the internet—they've seen more oral broadcast material than I could in several lifetimes. When it comes to "how to write," there's no reason they should be worse than me.
So I just threw the article at GPT and asked it to convert it into an oral broadcast. The result was unexpectedly bad.
It would repackage the original article, add a bunch of self-Q&A and explanations of terms, break sentences into tiny pieces, and then throw in a few "Hello everyone"s.
Even without knowing how to write an oral broadcast script, I could tell that reading such rambling, choppy stuff out loud would be nothing but awkwardness.
You need to know what's wrong.
Several attempts at letting GPT generate directly gave poor results, so I figured I should build a Skill to constrain the model's thinking process. My initial requirements were only that it couldn't arbitrarily change things or add new viewpoints, so as not to turn the article into another knowledge-blogger's rubbish copy.
Sure enough, the model behaved this time, but new problems emerged. It just broke long sentences into shorter ones, converted Arabic numerals to Chinese numerals, and changed English names to Chinese names. The article barely changed—reading it felt like reciting my official account article verbatim.
At this point, I noticed a very prominent issue: reading and listening are two different experiences. In a reading context, we can build rhythm through typography, and authors can leave several foreshadowing clues at the beginning, only revealing the conclusion at the end.
But audio-visual content is a single linear flow—information can only accumulate from front to back. The foreshadowing and clues that the beginning can carry are very limited, so expression needs to reduce the listener's cognitive load, ideally allowing the listener to absorb everything without thinking.
So the problem with this Skill suddenly became very specific. I needed it to preserve the basic facts, my viewpoints, and judgments from the article, but in terms of information output order, it needed to be adjusted so that listeners could follow along without going back.
After iterating several versions and achieving consistent output across different models through a clear execution process, I finally got a version I considered usable.
The value of a Skill isn't in the Skill itself.
From my experience, you can see that even if you know nothing, forcing together a Skill isn't that hard. Many people nowadays like to claim their Skills are the essence of years of industry experience—hyping a set of implementation principles and operational processes as a treasure of civilization is self-deception.
Most people's professional abilities aren't unique enough to be exclusive to them. How to write sales emails, how to interview users, how to analyze financial reports, how to write ad copy, how to create product plans—there are already many excellent people on the internet who have shared these. The SOTA models have already absorbed large amounts of this kind of content during training.
So when I made the oral broadcast Skill, I never thought about becoming an oral broadcast expert and then teaching that knowledge back to the model. In many fields with abundant public knowledge, there's no need to spend a lot of time teaching the model the "how-to" that it already knows.
What actually needs to be supplied to the model is what result we truly want.
Think back to the last time you used AI to do something—how did you command the AI? After the AI produced the result, did you specifically point out where the problems were, or did you just reply with "optimize it again"?
What truly needs to be built is the ability to define what "doing well" means, and the awareness of constructing an evaluation system.
How do we build an evaluation system?
Looking back at my process of making this Skill, I actually never solved the problem of "how to write an oral broadcast script." Throughout the interaction with GPT, I kept pointing out errors and constantly adding acceptance criteria.
Let the model do it first, and after seeing unsatisfying results, keep asking yourself what exactly you can't accept. Once you clearly articulate the vague dissatisfaction, the evaluation criteria begin to appear.
When there are enough criteria, rules gradually form. When rules have a sequential relationship, processes and standards naturally emerge, and eventually, they become a Skill.
You don't need to master a methodology in advance—you just need to keep converging during the process, building it through dynamic collaboration with the model.
This oral broadcast script conversion Skill grew out of that process.
The model is responsible for finding the path; I'm responsible for judging whether that path is getting closer to the result I want.
So if there's anything of my own in this Skill, I think it's not the knowledge about oral broadcasting, but my evaluation of the results.
Forcing together a Skill isn't that hard—you just need to know what you can't accept, and the rest is the result you want.