Jun Sasaki / Log

2026-07-20

A week where GPT updates increased what I can do.

Weekly Metrics 2026-W30

Last 7d 2026-07-13 to 2026-07-19 vs Previous 7d 2026-07-06 to 2026-07-12

Channel views Total watch time (min) AVD Rt2112 3 ▼50% 47 ▲840% 15:48 StudyRiver 18,256 ▲18% 1,974 ▲25% 0:17 StudyGym 3,341 ▼69% 912 ▼36% 1:01 StudyJaEn 10,241 671 0:16 MazeMarathon 1,903 155 0:17

Reference (28d cumulative)

  • Rt2112 28d cumulative total watch time: 122 minutes

  • StudyRiver 28d cumulative total watch time: 4,962 minutes

  • StudyGym 28d cumulative total watch time: 2,416 minutes

  • StudyJaEn 28d cumulative total watch time: 671 minutes

  • MazeMarathon 28d cumulative total watch time: 155 minutes

What I did this week

  • Shoten.me: Revised operations because the expectations for weekly inventory drafts and daily store openings were misaligned. Selected and verified 4 themes and 48 books for this week, and published 4 shelves after human review. Assuming a routine of 1 store per day starting this week.

  • Rt2112: Completed preparations for August Podcast and other distributions. Changed manga Shorts to 2x2 grid motion comics, and completed production and publication preparation for the next 28 volumes. Due to capacity, migrated distribution jobs from laptop to desktop PC and verified posting status to avoid double posting. Moved various materials (audio, images, etc.) to an external HDD, and completed ledger reconciliation and restoration verification. New operations started this week.

  • StudyRiver: Migrated to HDD from a storage operations perspective. Following the rejection by TikTok Developer, performed branding revisions on pages under eidendo.co.jp. Content continues to be distributed.

  • StudyGym: Reorganized the channel to specialize in Brain Training and separated the maze series into a different channel (Maze marathon).

  • StudyJaEn: Newly established. Completed and listed 4 English writing drills and free samples, and started automated operation of the video channel. 5 Shorts per day, 1 Long per day. Replaced the 4 videos where answer display errors occurred in the long-form videos, and re-checked all 4,000 sentences in the PDF.

  • System Trading: Changed to a system where daily verification is performed independently by Codex and Claude, with human comparison and arbitration. Discovered and restored the stopped tick collection for this week, and am waiting for market-hour confirmation next week. Various verifications show good results, but continuing due to small sample size.

  • For the Rt2112 Podcast, I began designing a structure that separates distribution origins to reduce manual work for each distribution destination while leveraging multilingual capabilities on the YouTube side.

Verification

  • For Shoten.me, to maintain the cumulative approach, I utilized MCP to create a routine. As a result, interesting bookstores are being mass-produced daily. Integration with Threads (SNS) is the next point.

  • For Rt2112, while only Shorts were generating reach, the median AVD for videos of about 30 seconds was about 8 seconds. I judged that the bottleneck was not the distribution volume but the presentation at the beginning, so I changed to motion comics. Since long-form views are almost zero, short-form (manga) is effectively the first step toward a breakthrough. I will continue to keep the priority low while maintaining production and verification. This week is publishing week.

  • StudyJaEn hit its highest initial growth with 26 registrations in one week. The first Shorts video reached 1,605 views in just under 24 hours, showing an initial response that suggests Shorts can be a primary battlefield. On the other hand, a long-form layout error occurred, so I have made full inspections before mass production a permanent practice. Connecting TikTok and Instagram is the task for this week.

  • Additionally, including this log, I am currently experimenting with operational routines regarding where I should intervene as feedback.

Hypothesis

  • The current bottleneck has shifted from production volume to a 'mechanism for observing whether automated operations have achieved the expected results.' Aside from corporate projects where I prefer to keep my feedback aside, I have moved the Study series into an AI loop improvement operation. Currently, weekly reports are uploaded from each project directory on weekends, and the AI itself holds a Sunday meeting and provides feedback. (I only intervene if I have objections to this feedback.)

  • As a result, it seems there is no issue with increasing production volume (i.e., mass-producing new Study series), so I will raise the baseline through horizontal expansion.

Next week's move

  • Complete the publishing work and the connection to automated SNS posting to advance the lifelong learning manufacturing industry.

The optimal solution for storage

The PC storage issue has finally become a bottleneck.
My personal bottleneck is the WAV data for audio and BGM. I make sure to delete video files once they are uploaded.
This week, in addition to the work mentioned above, I've been overhauling my BGM channels, creating a system to consolidate the workspaces of three channels into one.
When doing so, I create a "bgm_factory" directory on both my local (SSD) and HDD, and I make sure to move files to the local drive when generating music.

Originally, I thought about creating a system to pull from Google Drive, but data transfer takes time regardless. Since I run my iMac 24/7 for the Study series, I decided to operate it on the desktop in the same way.
I considered buying an external SSD, but I had a 2TB HDD lying around, so I figured that would be fine. Even if it's a bit slow, it's work done while I'm sleeping, so I don't mind.

About the BGM channels

I operate 1) Sleep2112, 2) Lo-fi2112, and 3) Pre Hours Vibes, each with around 10 subscribers. Channels 1 and 2 were originally created to reuse BGM from Rt2112, and channel 3 was created because I had leftover Suno credits. Honestly, while these three channels were created out of a sense of wastefulness and ended up producing nothing, I also wanted to make music as I pleased, so this time I'm overhauling them to put them on an automated track.
First, channel 2 is for making whatever I want. Channel 3 will be put on the automated track, and channel 1 will be linked to it as well.

I am currently stocking a large amount of Suno BGM on my HDD (downloading is very tedious).
I am sorting these by categorizing them by music BPM and theme. I will then randomly retrieve the necessary duration from the categorized songs to generate videos.
For those options, I have created skill sets such as timers, Pomodoro, and equalizer effects.
Channel 3 originally had no philosophy, so I am tailoring it to demand by making it a time-based work BGM channel. For example, I've created skill sets where you can choose options like 25/5 or 50/20 for Pomodoro.
I want to reach a style where I create one BGM video once a day during the late-night hours.

Differentiating between YouTube Shorts and long-form videos

In the end, you never know what will hit until you release the video. That's what it comes down to.
I split the Study Gym channel, which was a mix of spot-the-difference and brain training, into "Brain Training" and "Maze" channels. The reason for this is that Shorts were gaining viewers from the start for "Maze," while long-form videos were doing the same for "Brain Training," and since Shorts were popular in Japan and long-form in Asia, I prioritized the long-form content.
While I thought long-form was mostly India and Pakistan, it also includes viewers from Brazil.
The math channel StudyRiver is exclusively for Shorts, with 40% of the audience being seniors worldwide, including viewers as far away as Morocco and Peru. Study Lang was originally intended for Japan, so 90% of the audience is in Japan, with a wide range of ages.

I set some expectations for personas, but you never know until you try.
Since I'm 40, when I watch, noise runs through and is distributed to those areas. Study Lang Ja En was created with this in mind. In fact, 35% of the audience is 25-35 years old, which is exactly the working adult demographic that has English learning challenges.
For both Study River and Gym, I intentionally eliminated my own noise to reach a global audience, and as a result, the initial performance has become stable.

However, one thing I can say is that without Shorts, YouTube is almost an impossible game for long-form content.
Automated operations will fail 90% of the time unless they are built on the premise of running Shorts.

Running two AIs, with humans only acting as arbitrators

Since GPT Sol 5.6 is more capable than I thought, I've decided to operate automated routine systems in conjunction with Codex. Although there are still some suspicious behaviors, I am hedging risk by having one "opinion board" that acts as a human would.
I always place "docs" and "reference" directories in each folder; "docs" is for internal specifications, and "reference" is for external specifications.
Brainstorming sessions like "wouldn't it be efficient to incorporate this from deep research or internet information?" are generally stored in "reference," and then promoted to "docs" once approved.
It feels like I'm having Claude Code and Codex brainstorm these things.

Right now I'm typing commands one by one, but while I think it might be okay to design a loop for this, I'm not doing such deep work yet, so it's not worth implementing and is currently being done irregularly.

Summary

Last week, there were good GPT updates and a credit festival, so it felt like I implemented every move I could. The newly established StudyLang JaEn is growing more than expected, so it might be worth growing from Japan after all.
Personally, I call this "lifelong learning manufacturing," and since there's a limit to what one person can do, I will continue to lay down automated tracks. However, when thinking about it as a long-term business, I think there are unexpectedly many ways it could develop. I'm not at that turn yet, so I've decided not to think about it.

Advertisement