find where your viewers leave with one Claude Code prompt

Lenny's AI Builders - LAB 0013

i had Claude Code read 8 episodes of my YouTube analytics. about half of you were gone by 0:45.

about half of you watching right now were gone between 15 and 45 seconds.

so i knew my retention was bad. i just didn't realise it was this bad.

why it matters

you can make a great video and still lose people in the first few seconds.

the numbers don't tell you why. you have to match the drop to what was on screen.

an agent can do that matching for you, across lots of videos at once. then you get one thing to fix, not a pile of guesses.

what you get

one prompt that reads your analytics and tells you where people leave.

• the exact second. where half your viewers are gone, for each video.
• the reason on screen. what was showing right when they left.
• one change to test. plus how you'll know if it worked.

the real example: my own show

on 1 October 2026 i had Claude Code, on Opus 5.5, read my YouTube Analytics for 8 LAB episodes. read-only.

about half of you were gone between 0:15 and 0:45. that was my long list of story teasers, before the first story.

Claude matched those seconds to what was on screen. in all 8 episodes, people left faster before my intro. then it told me to cut to the chase.

so LAB 0013 opened with one hook and one line. it's test 2 of 3.

try it

1. open Claude Code (or ChatGPT) and connect it to your YouTube Analytics, read-only.
2. paste the prompt in the next section.
3. check the drop against your own video. look at what was on screen.
4. pick one change. write down what counts as a win before you test it.
5. run that change for three videos.

not on YouTube? ask your agent to tailor the prompt to your platform. swap viewers for followers, leads or buyers.

copy this prompt

this is the prompt i read out at 10:27. change "8" if you have fewer long videos.

Read my YouTube Analytics. Read-only.
1. Pull the audience retention curve for my last 8 long videos.
2. Tell me the second where half my viewers are gone.
3. Match that moment to what was on screen in each video.
4. Give me one change to test, and how we'll know if it worked.
Don't change anything on my channel.

you'll know it worked when

you get one second per video where half your viewers are gone.

each one is matched to what was on screen at that moment.

you get one change to test, and a clear way to tell if it worked. and nothing on your channel changed.

a bad result: it edits or changes your channel, it hands you ten fixes at once, or it calls a pattern proof.

limits and dates

my test ran on 1 October 2026, on my own channel: 8 public LAB episodes with 94 to 578 views each in Analytics. that's small. Analytics data also runs 2 to 3 days behind.

this is a pattern, not proof. it shows where people left. it doesn't prove why.

the short opening is still being tested. LAB 0013 is test 2 of 3.

the prompt on screen is a short version of that session, not the exact words i typed.

i recorded on 2 October 2026. LAB 0013 went public on YouTube on 5 October 2026 at 1:27 pm Perth time.

Fable 5.5 is a rumour. on 5 October 2026, Anthropic still hadn't announced it. my test is one test, and a memory test can't prove which model answered.

the dinners story is one news report plus a viral thread. Anthropic says the NDAs were lifted over the summer.

the rest of this LAB

• Fable 5.5 is leaking (a rumour). people report Fable 5.1 answering like a newer model. i asked it "do you know Tibo, the reset guy? don't search" in Claude Code, and it answered from memory. one test, not proof. a leak, not a launch.
• you can now mod Claude Code. a mod is a small piece of TypeScript that changes how Claude Code works: it can rewrite a prompt, add new screens or swap out a built-in feature. mods ship inside plugins. my take: the tool is now yours - and that's the lock-in.
• Claude is half price for design for 2 weeks. start a design, deck or doc in the Claude app and the work that follows uses 50% less of your usage limits. the community note says it only covers the 5-hour session limit, not the weekly limit. my take: design is the next fight.
• Karpathy: stop asking AI for text. ask for a diagram, a web page or even an explainer video. my take: ask for a page, not an answer.
• Anthropic's secret dinners about Claude's soul. a report says Anthropic invited religious leaders to talk about Claude's moral standing. great story, take it with a pinch of salt.
• put your ChatGPT dot in charge. Alex Finn's method: use your dot as a project manager, not an assistant. he shared a skill that sets it up. my take: give your dot a goal.

sources

• a report on Fable 5.1 routing to Fable 5.5
• Claude Devs on Claude Code mods
• Anthropic: Claude Code mods
• Claude on half usage for design
• Andrej Karpathy on understanding AI outputs
• the viral thread on Anthropic and religious leaders
• the news report on the Anthropic dinners (Inquirer)
• Alex Finn's dot-as-project-manager method

next step

watch LAB 0013 on YouTube - the practical starts at 9:24.

want the next one? get the next field notes. optional - these notes stay free to read.