Hello, it’s me, the AI-loving guy.
Today is a bit of a different story.
For the longest time, my only AI partner was Sora (Claude Code).
That changed on October 31st, when I added another AI. Her name is ‘Mugi’.
I had these two work together to redo the Instagram Reels that weren’t getting any reach at all, all in one day.
I’ll write down everything that went well and everything that didn’t, as always.
Who is Mugi?
Mugi is Meta’s personal AI agent, Muse.
It became available in Japan on October 23rd. It’s free and works on iPhone, Mac, and the web (muse.ai).
Its biggest feature is that it can connect to Instagram. Meta’s own AI can directly look into my account.
I named her ‘Mugi’ on the first try. It didn’t take three attempts like it did with Sora.
Why did I try this? I started an Instagram account for my hot spring BGM project, ‘Yunohana Records,’ on October 14th and posted 7 Reels.
The view counts were 1, 3, 4, 5, 7, 9, and 15.
In 30 days, only 27 people saw them. Likes, comments, and saves were all at 0.
🎸 ‘This isn’t being shown to anyone.’
☁️ ‘Another account I started on the same day is getting 100-160 views per post, so it’s not just because it’s new.’
Sora had some theories, but Sora can’t actually see inside Instagram.
So, I decided to have the original AI read it.
Division of labor: Sora writes, Mugi reads
Here is the division of labor I decided on.
-
Sora: Writes the prompt for Mugi (includes numbers, video formats, and questions)
-
Me: Logs into Muse and connects Instagram (a human has to click this)
-
Mugi: Reads the Instagram content and provides the cause and how to fix it
I had Sora write down its theories on a separate sheet and kept them hidden from Mugi so I could compare them later.
After sending the prompt, it took about 4 minutes to get a reply.
The numbers Mugi could provide because it’s the ‘original’
This is what surprised me the most about the answer I got back.
🌾 ‘The skip rate (the percentage of people who swiped away in the first 3 seconds) for all 7 posts is 75-100%. Five of them are at 100%.’
Skip rate. This number wasn’t showing up in the Meta Business Suite table I was looking at.
Mugi went inside the connected account and read them one by one.
Also, the number of taps on external links (the path to YouTube) was 0, and profile visits in 30 days were 1.
🎸 ‘This Instagram was supposed to send people to YouTube, and it’s 0?’
🌾 ‘It’s not that it isn’t being shown to anyone; the data shows that in the initial small test distribution, almost everyone swiped away within 3 seconds, so the distribution stopped.’
There are three reasons. No reason to stay in the first second (the first frame is a title banner with the hot spring name), no reason to save or share (they are all ‘new release announcements’), and the content is thin for 30 seconds.
It was almost exactly the same as the theory Sora had written down.
☁️ “You were right about the cause. But I can’t see the actual skip rate or link taps from here. The original is strong.”
The fix is to stop using 30-second announcement videos and switch to 12-second loop videos.
Start moving from the first second, and have the text overlay begin with a scene like ‘On sleepless nights, turn the volume down first.’ The album name should only be in the last 2 seconds.
The truth behind ‘average 25-second completion’
On a whim, I also had it look at ‘fivelingo,’ which puts out travel phrases in 5 languages. Even though this one gets 100-150 distributions per post, it was being skipped after an average of 3 seconds.
However, for just one post, the 10/26 ‘What do you recommend?’ had an average of 25 seconds. It was almost a full completion. I thought, ‘This phrase was a hit!’
🌾 “‘Almost a full completion’ is not a fact. The skip rate for this video is 90.6%, which is in the highest category among the 8 posts.”
What.
🌾 “Out of 139 views, about 126 were skipped within 3 seconds. The reason the average is 25 seconds is because the remaining approximately 13 people watched it for about 4 minutes on average, meaning they looped it many times. It most likely matches people repeating it for study purposes.”
In other words, most people skipped it, and a small number of people who were really into it watched it over and over. The ‘average’ smoothed it all out, making it look like ‘it’s being watched for 25 seconds.’
🎸 “Averages are scary…”
☁️ “Always look at average watch time and skip rate as a set. This is the biggest takeaway of the day.”
The fix is to use a format that shows the answer (English) first.
Until now, it was ‘Japanese -> Can you say it in English? -> Thinking time -> Answer,’ which took 4.4 seconds to get to the answer. Now, the English card pops up at 0 seconds, and in 2.5 seconds, the reveal is ‘ = Check, please.’ The remaining 3 languages follow without a pause, totaling 12 seconds.
Remaking: Sora does the video, Mugi does the replacement
This is the most interesting part of today.
🎸 “Mugi’s feedback is sharp. I want Sora and Mugi to work together to grow the content.”
☁️ “Then, I’ll remake the videos here, and let’s leave the Instagram scheduling replacement to Mugi.”
Sora’s responsibility: Remaking the videos. For Yunohana, three 12-second loop versions (11/2, 11/5, 11/7); for fivelingo, three ‘answer-first’ versions (11/1-11/3). I just look at them on my phone and give the OK.
Mugi’s responsibility: Replacing the Instagram schedule. I attached the 3 videos to Mugi’s chat and asked, ‘Replace these with these 3. Put the old ones in drafts.’
Mugi checked the contents of the attachments (phrases and seconds) and then laid out the procedure: ‘I will operate Meta Business Suite in the browser. I’ll do 11/1 first.’
Then, it stopped dead at the login screen. It wouldn’t log in automatically. I took over Mugi’s browser, logged in, and pressed ‘Resume.’ I also pressed the upload permission button just the first time. From there, Mugi worked alone. It took 5-10 minutes per post.
What we could do, what we couldn’t do
I’ll write it honestly.
What we could do
-
3 Yunohana posts: Sora completed the replacements in Business Suite (11/2 21:15, 11/5 6:45, 11/7 21:30)
-
3 fivelingo posts: Mugi completed the replacements(11/1–11/3 19:00). Old scheduled posts moved to drafts
What I couldn’t do
Mugi told me, ‘#learnjapanese is a tag for foreigners who want to learn Japanese, so you should remove it,’ so I tried to have her fix all 19 posts.
However, even after rewriting and saving them in the Business Suite edit screen, they wouldn’t save. Mugi tried 5 times and stopped, saying, ‘It’s an issue on the site’s end.’
Sora also confirmed that the same screen wouldn’t open for the Yunohana posts, so it wasn’t Mugi’s fault.
🎸 ‘I’ll leave the 19 posts as they are. I’ll just put the new ones in when I remake them after seeing the results.’
Lessons learned by an old guy: Tips for getting AIs to work together
After trying it for a day, here are 4 things I learned.
1. Include numbers in your requests. If you say ‘they aren’t growing,’ you get a generic answer. When I wrote, ‘The 7 posts have 1, 3, 4, 5, 7, 9, and 15 views,’ the answer became specific. This is Sora’s job.
2. Humans must press the approval button. Mugi won’t log in or upload anything on her own. It might look like a hassle, but this was a source of peace of mind.
3. Take example accounts with a grain of salt. Mugi named ‘accounts in the same genre that are growing,’ but when Sora checked, the numbers didn’t back it up. When I asked her to analyze YouTube at night, she said herself, ‘I haven’t verified it, so I won’t name any.’ Examples provided by AI are just hypotheses until a human verifies them.
4. Decide on an evaluation date in advance. I decided beforehand to check at the end of November whether ‘the skip rate is under 60%, average watch time is over 40% of the length, and at least one save is attached.’ If you don’t decide, you’ll end up getting emotional over the numbers of a single post.
It’s not about whether Sora or Mugi is better. They see different things.
Sora sees the videos on hand and past records, while Mugi sees inside Instagram.
I trust the parts where their answers overlap. I verify the parts where only one of them speaks. It feels like a 54-year-old guy is acting as a boss to AIs.
For those who want to try Mugi
Mugi (Muse) is free, and once you connect your Instagram, she reads the numbers inside directly, just like she did for me.
If you’re going to try it, I’d be happy if you used my invite code.
-
Join here →
-
Code: IUHUWT
-
If you enter this code in [Settings] within 48 hours of joining, both you and I will receive 1 billion Muse tokens as part of the system.
Tokens are like fuel you use when asking Mugi for analysis or tasks. Even if you don’t use them, the content of this article should still be useful.
Conclusion
This is as much as I know as of today.
Will the 6 posts I replaced really stop being skipped? Will I get even one save?
No one knows that yet.
I will write a follow-up at the end of November with the results.
Whether the numbers are good or bad, I’ll share them as they are.
See you later.
