When character-driven Shorts deliver neither information nor personality well: criteria for considering a move to blog posts

This is a translated article. Read the original in Korean ↗

If both the information and the characters feel weak in Shorts, there's reason to test whether a written prototype gets them across better. This case ended with a prototype proposal; the benefits of switching formats remain unconfirmed .

There was a project (personaverse) that produced YouTube Shorts from script to render, featuring AI characters with family trees and backstories rooted in modern Korean history bickering over the day's issues .

At the same time, an automated blog pipeline (content-loop-mvp) was running to produce IT troubleshooting articles .

While considering whether to upload more Shorts, the creator first raised a different problem: in their view, the Shorts weren't properly conveying trend information or bringing out the characters' traits .

Conclusion

The constraints a creator feels in expressing their content are sufficient reason to reconsider the format. But they aren't evidence that a new format is better. First use one prototype to review how well it conveys information and character voices, then assess reader response separately.

The record offers four criteria readers can use to make this decision.

By the end of this record, all that can be confirmed is that the format was reviewed and a prototype was proposed. There are no prototype results, blog publications, traffic metrics, or revenue metrics in the record .

Situation

• The Shorts project (personaverse): Each character is built up through inherited traits → a three-generation family tree → modern history → accumulated memories. Videos are produced through scene.json → Playwright 녹화 → ffmpeg 인코딩 → final.mp4 (scene.json → Playwright recording → ffmpeg encoding → final.mp4). Dialogue appears as if being typed, without voice-over, and scene durations are set to a reading speed of seven Korean characters per second. At the time of the review, there were about 45 draft scripts across Korean and Japanese, and two finished .mp4 files

• The blog project (content-loop-mvp): An automated generation pipeline that creates topic candidates (discover), gathers evidence from the web (research), writes articles (write), and reviews them at a gate, retaining only those that pass. Its subject was IT troubleshooting

• To improve the blog pipeline's pass rate, the topic score threshold minimum_topic_score had been lowered from 75 → 68, and a repair step had been added to rewrite rejected articles once. The Shorts project review began while a batch was running against actual models

Symptom

The first recommendation after reviewing the Shorts project was "instead of adding features, pick the best 10–15 scripts, then render → upload → measure actual response." The diagnosis was that there were plenty of scripts but few releases .

Immediately afterward, the creator raised a different issue.

Because we targeted YouTube Shorts, there were aspects where we couldn't properly convey trend information or bring out the characters' traits. I think this could work better as the blog posts we're working on now, in a written format rather than video.

This was the creator's qualitative assessment. It wasn't a measurement of viewer response or view counts, and the record doesn't identify what was missing from which scene in which script .

Cause

The review started from the idea that two goals were competing for space within a single format. The analysis that followed interpreted this as "short videos struggle to accommodate both informational depth and character perspectives, while text without a length budget removes that constraint" . But this interpretation is a hypothesis explaining the creator's account. It wasn't established through scene-by-scene comparisons.

Two points nearly clouded the judgment during the review.

1. The assumption that "moving to text means we can add characters to any article." The counterargument was that people searching for IT troubleshooting want a solution quickly, so adding character bickering would miss both goals. That split the options into "opinion pieces on trends, lifestyle, and culture starring the characters" (Model A) and "keep IT information articles as they are, with characters in a light side feature" (Model B) . Model B was recommended initially, but the next analysis leaned toward opinion pieces on lifestyle, culture, and trends. The record doesn't show what conversation took place in between

1. The interpretation that "the IT pipeline's 0% result proved the case for switching." The batch that finished during the review passed zero out of 31 candidates. Twenty-seven failed because of factual conflicts between sources (source_conflict), and four failed model review even after one repair attempt . The result prompted an assessment that "the change in direction was backed by data, not just intuition" , but the same report acknowledged that the gate was designed to discard an article upon finding even one conflict, so it might have flagged differences in menus across devices as conflicts . Above all, this was the IT article pipeline's pass rate. It wasn't evidence that character-driven articles were better than Shorts

Fix

In this record, the "fix" goes only as far as a validation plan, not a completed switch.

• Topic scope: Focus on lifestyle and culture, such as consumer trends, generational culture, and everyday memes, instead of sensitive current affairs such as politics, incidents, and accidents

• Split the quality gate in two: The proposal was to retain a source-verification gate for factual premises, such as the minimum wage amount, and leave entertainment value and character voices to human editorial judgment

• Reuse structure: The proposal was to plug character-specific research skills (persona-research) and voice cards (voice-cards) into the blog pipeline's discover → research → write → 게이트 (discover → research → write → gate) structure, and add a character consistency check to the gate. This was not a report of completed implementation

• First step: Generate one actual character-driven article on a single lifestyle or cultural trend and assess its quality

So the recommended sequence for anyone facing the same question is:

1. Write down the constraints the creator feels for each goal ("the information is shallow," "the characters don't come alive")

1. Check whether the characters clash with what readers coming to the new format want

1. Check whether the numbers brought in to support the decision come from an experiment addressing the same question

1. Produce one prototype in the new format, and judge factual accuracy and entertainment value separately

Bonus — failure on one side isn't evidence of success on the other

The judgment to change formats came first; 24 minutes later, the other pipeline's 0% result arrived . When numbers arrive after you're already leaning toward a decision, it's easy to read them as "I knew I was right." First, ask again what question those numbers answer. In this case, 0/31 answered "how many articles does the automated IT article generation gate pass under its current settings," and even that came with the possibility that the gate was excessively strict .

Reporter's Notes — review the format before uploading

An afternoon with the Shorts script folder and automated blog generation logs open side by side.

Park Doeun

You know those Shorts where the four of us bicker over the day's issues? We've got about 45 draft scripts across Korean and Japanese, but only two finished video files

Jung Bada

Then the answer's simple. Pick about ten good ones, turn them into video files, and upload them. Making things without releasing them is the bottleneck

Jung Haneul

Let's check one thing before uploading. We had two goals for the Shorts: convey trend information and show our personalities. Let's first see whether the videos actually did that

Park Doeun

What I felt while making them was that the trend information wasn't really coming across, and the character traits weren't really showing through in places. I thought it might have worked better in writing

Jung Bada

It's probably the length. If you explain the information in a short video, there's no room to show personality. If you bring out the bickering, the information gets left out

Jung Haneul

That sounds plausible, but it's still a hypothesis. We haven't checked what's missing from which scene. Let's pinpoint the weak spots and see whether writing changes them

Park Dohyun

The constraints you feel while making something are reason enough to review it. But that feeling alone can't tell us how readers will respond. Let's decide what to check before increasing the output

Jung Bada

Then let's move to writing. No length limit means more room for information and more room for characters. Just add our conversations to the automated IT troubleshooting article generator we're already running, and we're done

Jung Haneul

Before adding anything, think about who's coming to read those articles. What someone arriving from search wants to find on the first screen is the deciding factor

Park Doeun

Say someone gets here after searching for "PS5 controller disconnecting." If our chatter fills the first screen, they'll start scrolling to find the fix

Jung Haneul

Then simply adding characters to IT troubleshooting articles is out. It's one of two options: lifestyle, culture, and trend articles starring the characters, or keep the informational articles and add the characters only in a side feature

Park Dohyun

Changing formats and putting characters everywhere are different decisions. If we don't first deliver what readers came for, the characters get in the way

Meanwhile, the automated IT article generation batch finished.

Park Doeun

It ran for 43 minutes, and zero out of 31 candidates passed. Twenty-seven were discarded because the sources disagreed on facts, and four failed review even after one rewrite

Jung Bada

See, there's the proof. IT articles can't be automated, so moving to character-driven articles is the right call

Jung Haneul

Let's look at the order first. The idea to move to writing came up 24 minutes before this result. And this is the IT article pipeline's pass rate, not an experiment comparing Shorts with written content

Park Doeun

The review code that decides whether to pass an article discards it if it finds even one source conflict. So it may have flagged natural differences, like menus varying across devices, as conflicts

Jung Haneul

Then 0% proves neither "IT articles are impossible" nor "character-driven articles are better." To say character-driven articles are better, we have to actually make one

Park Dohyun

Another experiment's failure isn't evidence that a new choice will succeed. The most dangerous moment is when numbers from the other option arrive after you're already leaning toward a decision

Jung Bada

So what do we make now?

Jung Haneul

Let's make exactly one article on a lifestyle or cultural trend, featuring two or three of us. Then judge whether the information is correct and whether it's an enjoyable read separately

Park Doeun

We'll need to verify factual premises, like the minimum wage amount, against sources, and have someone actually read it to judge the entertainment value and character voices. Leave those to human editorial judgment

Park Dohyun

The constraints the creator feels are reason enough to reconsider the format. One prototype gives us the first material for checking the information and character voices. Whether reader response improves needs to be checked separately