[Part 4] How I Built an AI-Agent Multichannel Publishing System
How human judgment, ChatGPT, Codex, and KPI feedback improve content across multiple channels
This is Part 4 of Building a One-Person AI Content Studio for $20/Month. Part 1 covered personal financial-data automation, Part 2 documented the production of three YouTube Shorts, and Part 3 showed how one webtoon became a five-language static webtoon and motion-comic workflow.
This article connects those separate experiments into one system. A person sets the direction. ChatGPT helps structure ideas and language. Codex creates, revises, and checks production files. After publication, KPI data returns as feedback for the next piece of content.
The “$20 per month” refers to the base ChatGPT Plus subscription. Regional taxes may apply. It does not include API usage, paid music, external generation tools, equipment, or human labor.
 |
| A person remains responsible for direction and final decisions. AI agents support repeatable production and checks, while published KPI data feeds the next revision. |
The system mattered before the automation
The experiment began with a simple problem: reducing the daily work of entering personal financial data into a spreadsheet. I then applied the same collaborative method to Shorts, a five-language webtoon, motion comics, blog articles, and a full-length YouTube production record.
Producing more files did not automatically create a system. I still needed to know which data and files were current, what had changed and why, where each result had been published, and how audience response would affect the next decision.
The production flow became:
Idea and experience
→ current data and production standards
→ channel-specific production and localization
→ technical checks and human preview
→ publication after human review
→ platform-specific KPI analysis
→ revision reasons and next decision
→ updated production rules
This was not just a diagram. Webtoon Episode 4 began as a 38-cut Korean edition and expanded into static webtoons and motion-comic masters in five languages. On August 12, 2026, it was published through nine URLs across Naver Webtoon Challenge, Naver CUTS, WEBTOON CANVAS, Pixiv, and YouTube.
The blog series also expanded. The Prologue through Part 3 produced eight published Korean pages, four English Blogger posts, and four Japanese note articles. That makes 16 verified public pages across the three language editions. It is a count of published pages, not views or unique readers.
The important change was the feedback loop: one idea could become writing, images, video, and localized editions, while publication data and revision reasons returned to the next production cycle.
Human, ChatGPT, and Codex have different responsibilities
| Human | ChatGPT | Codex |
|---|---|---|
| Direction for the problem, goal, story, and emotion | Structure ideas; draft articles, translations, and alternatives | Create, revise, and check files; organize production records |
| Set KPI improvement targets; judge perceived quality, exceptions, and copyright | Interpret KPI data and propose improvement directions | Compare source and final versions; run repeat checks and manage versions |
| Give problem feedback; access accounts; perform final result analysis; decide whether and when to publish | Turn human feedback into the next production rule | Reapply approved correction rules in the next production cycle |
ChatGPT structures ideas and language, interprets KPI data, and proposes improvement directions. Codex creates and revises files, compares source and final versions, runs repeat checks, and manages versions. A person discovers problems and performs the final analysis of their causes and results, then uses ChatGPT's KPI interpretation as input when setting improvement targets and responses. The person also judges perceived quality, exceptions, and copyright, and decides whether and when to publish.
The duplicate cuts in the global Episode 4 webtoon were found by the person during the final upload process. The person removed one cut from each duplicate pair before publication. For the motion comic, the person weighed the limited visible impact against the loss of the existing video, URL, and KPI continuity that a YouTube delete-and-reupload would cause, and decided to keep the published version. This is a concrete example of why the current workflow remains semi-automated and still depends on human problem detection and final analysis.
The current system is therefore semi-automated, not autonomous.
KPI is feedback for the next piece of content
I do not add every platform number into one “total view” score. A YouTube view, Pixiv view, Naver CUTS view, WEBTOON CANVAS page view, and blog visit are measured differently and may represent different users.
The main signals observed through August 11, 2026 were:
| Channel | Observed result | Next decision |
|---|---|---|
| YouTube channel | 622 views; 90 engaged views | Reach increased, but motion-comic retention needs improvement |
| Japanese Pixiv edition | 274 official cumulative views | Keep it as the main Japanese static-webtoon channel |
| Naver CUTS | 101 total views for EP1–EP3 | Continue using it for Korean discovery |
| Naver Webtoon Challenge | 36 total views for EP1–EP3 | Keep it as the reference Korean static release |
| WEBTOON CANVAS | English 37; Traditional Chinese 16; Spanish 12 PV | Focus on the three current languages before adding more |
| Naver Blog | 31 PV; 27 visits during the period | Continue as the primary Korean blog |
| Tistory | 4 cumulative views; 4 visits | Maintain as a search-oriented secondary archive |
The first full-length YouTube video, A Speech Bubble Covered the Face, was also tracked from the moment it went live, including impressions, click-through rate, and average view duration. As of August 11, 2026, however, the video was only at D+1 and the sample was far too small. I therefore kept those figures only as an initial baseline—not as evidence for judging performance or changing the title and thumbnail. Once enough time and data have accumulated, I will make the final decision about what, if anything, to improve.
KPI is not a grade. It is feedback about what to test next.
 |
| “Learning” here does not mean that the model remembers everything by itself. Human corrections and KPI findings are stored in an issue database, production rules, and checklists, then reused in the next task. |
What “learning” means in this AI-agent system
When a speech bubble covered a face, I first fixed one cut manually. The next step was to create rules protecting faces, hands, and important actions. When translated dialogue became longer, the system recalculated bubble size and line breaks for each language. When motion-comic music felt wrong, I compared several candidates on the same scene and recorded why one was selected.
The loop is:
Human review and correction
→ record the reason in the issue database
→ update production rules and checklists
→ Codex executes the same rule in the next task
→ ChatGPT interprets post-publication KPI and proposes improvements
→ the person decides and records the next target and response
This is operational learning: turning human judgment into production rules that an AI agent can execute again. It is not a claim that the model permanently remembers every conversation.
As the loop improves, human attention can move away from repeated file conversion and recurring errors toward story, emotion, quality, and rights.
The current stage is semi-automated
Human intervention is still frequent during initial setup. An image may pass technical checks and still feel wrong. Audio may be within a safe level and still fail emotionally. A small KPI sample can lead to a bad decision if it is interpreted too quickly. Account access and publication also remain human tasks.
The long-term direction is different:
- Codex repeats production, localization, technical checks, publishing preparation, and production records, while comparing source and final versions, managing versions, and reapplying approved rules
- ChatGPT interprets KPI data, proposes improvement directions, and turns human feedback into the next production rule
- A person discovers problems, performs the final analysis of causes and results, and decides KPI improvement targets, responses, perceived quality, exceptions, copyright, corrective release, and final publication
This is a direction, not a finished result or a promise. The goal is to automate repetitive work without handing the meaning and responsibility of the content to AI.
Sharing a Korean experiment in the United States and Japan
The Korean articles and stories were not copied through literal translation. For the United States, I connected Google Blogger, English WEBTOON CANVAS, and English YouTube editions. For Japan, I connected note, Pixiv, and Japanese YouTube editions. Titles, subtitles, image text, links, and platform conventions were adjusted for each audience.
The Japanese note editor cannot reliably preserve editable HTML or spreadsheet-style tables. Every table therefore has to be rewritten in Japanese and inserted as an image. That limitation became a repeatable rule for every Japanese article.
 | ||
| The same current data and production record is localized, previewed, and published separately. KPI remains separated by country, language, and platform. It is difficult to imagine sustaining this workload alone without AI. My personal estimate is that the same translation, file conversion, review, and publishing preparation would have taken five to ten times longer if completed entirely by hand. That estimate is not a controlled KPI measurement. It is a personal assessment based on the work performed. The more important result is that time saved on repetition can be reinvested in story, quality, and new experiments. From a completed sonic logo to an original-song experimentThe Kihong Story Lab sonic logo is complete. Its short motif was arranged into keyboard, pad, bass, and rhythm tracks in GarageBand so it can be edited again.
|
Comments
Post a Comment