I've run my Instagram for 15 years. Long enough to know two things hold no matter what the algorithm does: keep testing new formats, and work out afterward what actually did the job.

For most of those years the second half of that ran on instinct. A feel for pacing and hooks, built from posting almost every day. My reels go viral often. 

Views are not a goal for me. A community that shows up is.

That 9.1M reel was genuinely fun to shoot: I turned a restaurant tablecloth into a skirt to get past a dress code, and I expected it to take off. It brought 1,396 new followers, roughly 0.015% of everyone who watched. In absolute terms that's a good number. As a conversion, it's a failure: a quality viral reel converts at 0.02% to 0.05%.

The joke went far beyond the people I'm trying to reach. Millions watched, laughed, and never clicked through to find out that my actual content is about something else.

Going viral is nice. Going viral inside your niche is what I'm after. If growth is the goal, two numbers are worth watching: new followers and shares.

My target band is 150K to 300K views. Claude landed on that range after going through my account stats: it converts into new followers better than anything above it. Past that, the audience stops overlapping with mine.

🔴 So I'm running an experiment: my 15 years of instinct against an analytical partner in Claude. I hand Claude my own reels data, let it find the patterns I've been calling intuition, and see whether the next few videos beat my gut on the number that actually matters: new followers, not views. 

Here's the exact setup. Follow these steps to build the same one for your own profile or for a client's.

Claude Code and Apify (a marketplace that pulls public Instagram data). You don't need to know how either one works. You copy one prompt and answer a few questions.

Switch to Claude Code, create a new project and connect an empty folder on your computer, dedicated to this one.

Create an Apify account and connect it to Claude. Claude does the work, Apify feeds it your public Instagram data. When you run the prompt in the next step, Claude asks for your Apify key and wires everything up for you.

Download the full prompt here and send it as your first message. That's the whole installation.

Claude then works on its own: it researches Apify's current Instagram Actors and their real prices, builds the project structure, and writes its own operating manual (CLAUDE.md) so future sessions remember everything. At one point it asks you to paste in your Apify key. That's the only thing you add by hand. Just let it finish.

One note on setup. So Claude can read the audio and on-screen text in your reels, it may ask to install a couple of small free helper tools. You don't do anything yourself. When it asks, just approve it.

Take the time to answer properly instead of rushing or saying "I don't know." Claude will ask you about your niche, goals, audience, how many reels a week you can film, what you refuse to make. It'll propose competitor candidates, and you just confirm. It took me 20–30 minutes to gather all the information it requested and complete the interview.

Pro tip: screenshot your Instagram Insights (Audience + Content tabs) and drop the images into the chat. Public data can't see retention, saves, or follows-per-reel — your screenshots fill that gap.

Claude pulls the reels you've posted and ranks them by outlier score: how far each one's views sit above your own recent median. Which means it only knows views, so this is where those Insights screenshots earn their place: tell it which reels actually brought followers and shares, and it sorts on that instead.

It transcribes your best and worst, reads the comments on your top ones (the questions your audience is already asking), and hands back a report with what's working, a few real hypotheses, and a first batch of reel ideas with the hooks already written. Pulling 200 of my posts cost 46 cents. 

If you’re not posting regularly on Instagram yet, point it at a leader in your niche you look up to instead.

Now just talk to it:

"Write a script about [topic]"
"Ideas for the week"
"Break down competitor @X"
"Why didn't my last reel perform?"
"Weekly report"

Every script comes with 3 hook options, a timecoded shooting table, a caption, hashtags, a cover, and a hypothesis: what this reel is testing.

When you publish, tell it. It pulls the numbers at 24 hours, 72 hours, and a week, checks them against what it predicted, and updates what it knows works on your audience specifically. Skip this and it never finds out whether it was right. Do it, and it gets sharper every week.

I set it up for a weekly run.

I've said this before and I'll say it again: I never hand the creative part to any AI. Claude is genuinely good at generating ideas, but only I can validate them and decide if they fit my values, my mood, my style. A few weeks ago I wrote that AI has never once handed me a raw idea I said yes to, and not one of my reels started as copy-paste.

That's my call for a personal account, where the whole thing is me. Plenty of niches aren't like that. If you run a faceless page, a product account, or a channel where the value is the information and not the person, Claude can realistically carry 80–90% of it: the ideas, the scripts, the analysis. Mine just happens to be the kind where it can't.

I don’t expect magic, and I don’t expect this setup to turn into a full-time reel writer. But I believe in data and math, so we'll see if it works on my own account. And honestly, having some help sounds amazing, especially in busy stretches. I'll tell you how this experiment went in a month or so.

Marina 💜

P.S. Know someone who'll get value from this setup? Forward them this email.

Missed the last few posts? Here's what subscribers opened most:

→ What's inside my Claude 👀
https://siliconvalleygirl.beehiiv.com/p/what-s-inside-my-claude

→ 7 skills that make you irreplaceable
https://siliconvalleygirl.beehiiv.com/p/7-skills-that-make-you-irreplaceable