Teardown No. 118 · Video

15 to 20 Retention Points: How Steve Chou's Claude Code audit of his YouTube channel found the intro habit he had trained into it

Steve Chou runs the ecommerce site, podcast and YouTube channel My Wife Quit Her Job. He connected Claude Code to the YouTube Data API and the YouTube Analytics API, had it pull a year of data, and asked what was working and what was failing. It matched retention drop-offs to his transcripts and found that every video with a backstory beat between the hook and the first point lost 15 to 20 retention points. The cause was his own setup: he had trained Claude on every script he had ever written, flops included, so his style guide told it to open with backstory. He now trains it only on outlier winners.

Try this step

Connect two APIs. Claude Code needs both the YouTube Data API and the YouTube Analytics API. He asked Claude Code to walk him through the Google Cloud credentials. Expect about 30 minutes the first time, and paste a screenshot when a menu option it names doesn't exist.

The problem

About 60% of his traffic comes from subscribers, with weak search traffic, so the algorithm is not pushing his old catalog to new viewers. One video drove about 30% of the views over the period Claude looked at. And his scripts kept opening the same way. Co-host Toni Herrbach brought a similar question from a client channel that had been stuck under 10,000 subscribers for about 20 months.

The blueprint (as he describes it)

  1. Connect two APIs. Claude Code needs both the YouTube Data API and the YouTube Analytics API. He asked Claude Code to walk him through the Google Cloud credentials. Expect about 30 minutes the first time, and paste a screenshot when a menu option it names doesn’t exist.
  2. Pull a year and ask why. Claude pulled click-through rates, retention curves, traffic sources and transcripts, then answered which content worked, which flopped and why. The analysis took about five to ten minutes.
  3. Match the drop-offs to the words. With transcripts loaded, Claude pinpointed the stretch of each retention curve where viewers left. It was the backstory every time, so now he jumps from the hook straight to the first point.
  4. Ask the agent why it writes what it writes. When he asked why his scripts kept opening with backstory, Claude said that was how he talked in the scripts he uploaded, years of flops included. He told it to cut the habit.
  5. Train only on outliers. A winner is a video that does exceptionally well relative to the channel’s subscriber count. Those scripts keep his voice without the habits of the flops.
  6. Keep the human veto. Claude told him to drop AI search videos, which top out around 21,000 views. He is keeping them on purpose, because that is where ecommerce is going.
  7. Refresh before you upload. For old hits two or more years old, a new title and thumbnail can be worth more than a new video. He makes three thumbnails per video and asks several AI tools and the vidIQ grader to judge them.

Stated results

  • “every video with a backstory section between the hook and the first point lost 15 to 20 retention points.” The new scripts no longer open that way.
  • Claude’s plan: at least three more boring product videos in the next 90 days, drop TikTok Shop titles, refresh the top five lifetime videos.
  • His Alibaba alternatives video has almost 3 million views; one video got a million views in a single day.
  • A product checklist video that flopped on views produced about 50 times the email signups of a typical video, relative to its views.

Caveats

  • He sells courses and a free mini course, and the episode page promotes them.
  • The numbers are his channel analytics as he reports them. No retention gain from the fix is reported yet.
  • Claude Code is the agent; he reviews its advice and ignores some of it on purpose.
  • Next step is still an idea: feed his per-video tagged email signups into Claude alongside the YouTube data.
  • Handles: X @mywifequit, YouTube, Pinterest and Facebook are linked from his site. No Bluesky, Threads, Instagram or TikTok profile verified.

FAQ

How does Steve Chou use Claude Code on his YouTube channel?

He connected Claude Code to the YouTube Data API and the YouTube Analytics API, had it pull a year of click-through, retention, traffic and transcript data, and asked which videos worked, which flopped and why.

What did the analysis find?

Every video with a backstory beat between the hook and the first point lost 15 to 20 retention points. Claude found it by matching retention drop-offs to the transcripts.

Why did the AI keep writing that intro?

He had trained it on every script he ever wrote, including years of flops, so his style guide said he likes to open with context. He now trains it only on scripts from outlier videos.

How long does the setup take?

About 30 minutes for the two APIs the first time, and about five to ten minutes for the analysis itself.

Sources


Want agents like these working your distribution? Get It Built. Get new teardowns weekly: subscribe (double opt-in).

Get one teardown a week.

Same format, a new operator every issue. Double opt-in.