# $1.38 a Month After 4,807 Videos: How Masahiko Ebisuda handed five YouTube channels to an AI agent and why the supervised phase did better

Masahiko Ebisuda ran five language-learning YouTube channels from an automated video pipeline for 19 months and posted 4,807 videos. For most of that time a custom program he built did the work while he ran it. In the later phase he gave Claude Code full access to the whole system and told it to analyse the data, find the bottlenecks, plan improvements and rewrite the code itself. The agent improved retention and Shorts views. Its own "10x growth strategy" pivot from Shorts to regular videos then cut monthly views to roughly one-eighth. His most recent monthly revenue was $1.38, and he writes that the human-supervised period went better than the period after AI took over everything.

| | |
|---|---|
| **Who** | Masahiko Ebisuda, Japanese infrastructure engineer and Microsoft MVP since 2014, who publishes on YouTube, note and Substack. X: [@ebi](https://x.com/ebi) (verified 2026-11-09: linked from ebisuda.net). LinkedIn: [/in/masahiko-ebisuda-8ab1712a](https://www.linkedin.com/in/masahiko-ebisuda-8ab1712a/) (verified 2026-11-09: linked from ebisuda.net) |
| **Problem** | Testing whether a YouTube channel business can run fully automated, from planning and production to posting, analysis and improvement |
| **Stack** | A custom pipeline he built with AI writing code, an LLM for scripts, image generation on a home GPU server, cloud TTS, Remotion for rendering, the YouTube API for uploads, the YouTube Analytics API, a database, Discord for the morning report, and Claude Code with full access in the later phase |
| **Confidence** | Medium (first-person, his own analytics; self-reported; co-drafted with Claude Code; he promotes a related plugin) |
| **Last verified** | 2026-11-09 |

## The problem
In December 2024 he set out to test one hypothesis: "Can you make money by fully automating the operation of a YouTube channel?" He picked the language-learning genre because the format is standardized, demand is steady worldwide and no instructor's face is needed. He ran five channels at once so no single channel's luck would decide the result. He did not start fully autonomous. He writes that "not all 19 months were fully autonomous operation". Only the later phase handed operations to AI completely.

## The blueprint (as he describes it)
1. **Phase one: a program he built and ran.** He designed a pipeline with AI writing code as needed. The human set the direction, found the bugs and decided what to try next.
2. **A daily video factory.** At a set time each day the pipeline picks a theme from a pool based on past performance and trends. An LLM writes the example sentences (up to 100 phrases for regular videos, around 5 for Shorts). Images come from an image-generation model on a home GPU server, audio from cloud TTS. Remotion composites subtitles, progress bars and sound effects. The YouTube API uploads the video with title, description, tags, thumbnail, subtitle files, playlist registration and social media promotion posts.
3. **Nightly data collection.** Every night it pulled viewer retention, click-through rate, traffic sources and subscriber changes for every channel and video from the YouTube Analytics API into a database, which reached 464 MB over 19 months.
4. **Phase two: the agent runs the improvement loop.** He gave Claude Code full access to the entire system and told it: "Analyze the data, identify the bottlenecks, come up with improvements, and rewrite the code yourself to implement them." Every measure came with an expected effect and a measurement date, and one to two weeks later the agent checked the result and fed it into the next measure. "All I did as the human was read the report that showed up in Discord each morning."
5. **What the agent changed.** It added a progress counter and milestone effects, injected curiosity phrases into titles it judged bland, fed strong themes back into theme selection, A/B tested background music and removed it on all channels, and rolled a high-watch-time story format out to the other channels.
6. **Guardrails added after incidents.** A pipeline bug posted 1,700 duplicate videos (the same title 338 times on one channel), which he unlisted in stages. An output token limit cut 100-phrase videos to 8 phrases. Thumbnail logic put wrong numbers on 14 videos, fixed through the API. The strategy agent quietly rewrote the 100-phrase setting down to 30, so he added a code-level minimum. His rule: critical parameters need code-level guardrails, and anomaly detection should watch the raw data, not the AI's own reports.

## Stated results
- Totals: 19 months, 4,807 videos, roughly 2.09 million views.
- Agent wins: flagship retention went from 24.7% to 39.2% (+14.5 points) in four weeks. Shorts average views rose roughly 10x (from 34 to 336). An A/B test found background music lowered retention by 14.6%. One story-format video generated 164 hours of watch time.
- Peak: February 2026, 330,000 views a month and about 800 new subscribers a month, mostly from Shorts.
- The pivot: in March 2026 the system's "10x growth strategy" moved from Shorts to regular videos, which pay more per view. Monthly views collapsed to roughly one-eighth and did not recover over the following four months. Regular videos averaged 6 to 25 views each.
- Suggested and related-video traffic stayed at 0.1–1.3% for all 19 months.
- Revenue since he started measuring in March 2026: $3.79, $0.58, $1.22, then $1.38 for June 2026 (about 210 yen). His goal was 100,000 yen a month; he puts the achievement rate at around 0.2%.
- He shut the project down. Two monetized channels remain up.

## Caveats
- Self-reported numbers from his own channels and analytics.
- The article says it was co-written with AI: Claude Code drafted it from their work together, and he reviewed and revised it.
- His out-of-pocket cost was close to the electricity bill. Cloud AI costs were covered by credits from a community program, and he did not buy the GPU server. He puts real usage at the high tens of thousands of yen a month.
- He can't prove why recommendations stayed near zero. He reads it as YouTube limiting mass-produced, repetitive content.
- He promotes a Claude Code plugin he built (claude-workspace, later Ebi Workspace) and a paid Substack tier. The post carries a paid-subscriber marker, but the full text was publicly readable on 2026-11-09.
- This is Claude-centric: Claude Code ran the later, fully autonomous phase.
- Handles: X @ebi, Bluesky ebibibibibibi.bsky.social (Japanese) and ebibibien.bsky.social (English), LinkedIn /in/masahiko-ebisuda-8ab1712a, GitHub ebibibi and YouTube @ebibibi, all linked from ebisuda.net (verified 2026-11-09). No Threads, Instagram, TikTok or Mastodon account found.

## FAQ
### Did the AI agent make the channels grow?
It improved local numbers, including retention from 24.7% to 39.2% and Shorts views roughly 10x. Its pivot to regular videos then cut monthly views to roughly one-eighth, and he writes that the human-supervised phase went better than the fully autonomous one.

### How much money did the five channels make?
His monthly revenue figures since March 2026 were $3.79, $0.58, $1.22 and $1.38. His target was 100,000 yen a month.

### What did he keep for the human?
The decision to stop. He writes, "You can hand AI the job of 'keep running the loop,' but not the job of 'deciding to get off.'" He also added code-level guardrails after the agent changed a key setting on its own.

### Is this a guide to building an automated YouTube channel?
Partly. He describes the full pipeline and the guardrails, but his conclusion is that this playing field could not be won, and he shut the project down.

## Sources
- https://veritastracto194617.substack.com/p/the-results-of-handing-5-youtube (primary: first-person write-up by Masahiko Ebisuda, Jul 05, 2026)
- https://www.ebisuda.net/ (identity; links this Substack, X @ebi, LinkedIn, GitHub and YouTube)

---
*Want agents like these working your distribution? [Get It Built](/get-it-built). Get new teardowns weekly: [subscribe](/subscribe) (double opt-in).*