Build a viral idea detector with Claude

Why the view count lies to you

Instagram counts a view the moment a reel starts to play or replay, which is the platform's own wording in the reel insights help article. So a nine-second loop that one person watched six times contributes six views. Stack that behaviour across a big account and a mediocre post can carry a very impressive-looking number.

Then add the second distortion: you cannot tell from the outside whether a post was boosted with ad spend. Paid reach lands in the same view counter as organic reach. Both of these mean the raw number on a post tells you close to nothing about whether the idea was good, which matters a lot when you are about to spend a week rebuilding it.

The two numbers that actually mean something

Run them together and the logic is simple: the outlier score finds the posts that broke out, and the engagement check confirms the breakout was real. A post has to pass both before it earns a place in your idea bank.

What the maths looks like on real posts

1. 📊 Find the account's normal. Pull that account's last 20 posts and take the median view count. Say it comes out at 5,000. That is the baseline everything gets measured against.
2. 🔍 Score each post. A post with 50,000 views scores 50,000 ÷ 5,000 = 10x. A post with 8,000 views scores 1.6x, which is noise, not a signal.
3. 🌱 Check the engagement ratio. Work out (likes + comments) ÷ views for every post, then take that account's median ratio too. Say their normal is 4 likes and comments per 100 views, so 0.04.
4. 🚩 Compare the outlier's ratio to that normal. The 50,000-view post has 900 likes and 60 comments, so its ratio is 0.019, less than half the account's usual. Ten times the views, half the engagement quality. That is the shape of replay-inflated or boosted reach, and it gets dropped.
5. ✅ Keep the ones that pass both. A 12x post that also holds its normal engagement ratio is a genuine breakout. That is an idea worth your week.

What you need and what it costs

Wire it up, about 30 minutes

1. 🔑 Make an Apify account. Sign up at console.apify.com, then open Settings and go to Integrations to find your personal API token. Copy it somewhere safe. That token is what lets Claude run the scraper on your behalf.
2. 💻 Install Claude Code. On Mac or Linux, open Terminal and paste curl -fsSL https://claude.ai/install.sh | bash. On Windows PowerShell it is irm https://claude.ai/install.ps1 | iex. Both are straight from the setup docs. Then type claude to start it and sign in through your browser.
3. 📁 Make a project folder. Create a folder called idea-detector on your desktop, then in Terminal run cd ~/Desktop/idea-detector followed by claude. Everything the detector writes stays inside that one folder.
4. 📝 Write your watchlist. Make a plain text file called watchlist.txt in that folder with one Instagram handle per line, no @ symbol. This is the only file you will ever edit by hand.
5. 🔐 Store the token safely. Make a file called .env in the folder containing APIFY_TOKEN=your_token_here, and tell Claude to read the token from there. Never paste an API token into a chat message or into a file you might share.
6. 🏗️ Build the detector. Paste the build prompt below. Claude writes the script, runs it once on a couple of accounts to prove it works, then you point it at the full watchlist.

The build prompt

This is the whole spec. It is long on purpose, because every rule you leave out is a rule Claude invents for itself, and the whole point of this build is that the maths is not up for interpretation.

Build me an Instagram outlier detector in this folder. Read the Apify API token from the .env file in this folder, and read the account handles from watchlist.txt, one per line.

Write a script that calls the Apify actor apify/instagram-scraper to pull the 20 most recent posts for each handle in the watchlist, capturing at minimum: shortCode, url, ownerUsername, caption, timestamp, likesCount, commentsCount, videoPlayCount and videoViewCount. Cache each raw response to disk in a raw/ folder so re-running the scoring never costs me another scrape.

Then score every post with these exact rules. For each account, calculate the MEDIAN views of its 20 posts and the MEDIAN engagement ratio, where engagement ratio is (likesCount + commentsCount) divided by views. Exclude posts published in the last 48 hours from both medians and from the results, because they are still accumulating. Outlier score is a post's views divided by that account's median views. Keep only posts scoring 3.0 or higher. Then apply the engagement filter: drop any surviving post whose own engagement ratio is below 0.8 times its account's median ratio, because views count replays and can be paid for while likes and comments cannot, so a big gap means the reach was not earned by the idea.

Output the survivors to outliers.csv sorted by outlier score, highest first, with these columns: rank, account, outlier_score, engagement_health (the post's ratio divided by its account's median ratio), views, likes, comments, posted_date, url, and the first 150 characters of the caption. Print a one-line summary to the terminal telling me how many posts were scanned, how many cleared 3x, and how many were then dropped by the engagement filter.

Before you write any of this, tell me your plan in plain English and flag anything in my spec that will not work with the real Apify output schema. Then run it on the first two handles only so I can check the numbers before we spend credit on the full list.

Turn the CSV into an idea bank

A ranked spreadsheet is data, not ideas. This second prompt is where it becomes something you can actually film. Run it on the fresh CSV each time.

Read outliers.csv and turn the top 10 rows into an idea bank I can actually shoot from this week.

For each one, give me: the account and its outlier score, what the post is actually about in one sentence, and my read on WHY it broke out, choosing from a specific hook shape, a format, a timely topic, or a strong opinion, and say which. Then give me the transferable pattern in one line, meaning the part that would still work if the topic were completely different.

Then group the ten into themes and tell me if the same underlying pattern shows up more than once this week, because a pattern repeating across several accounts is a much stronger signal than one big post. Rank the themes by how well they fit my niche, which is [YOUR NICHE], and my audience, which is [WHO THEY ARE].

Finally, pick the three I should make and write me a hook line for each in my own angle, not a rewrite of theirs. If any of the ten are only interesting because of who posted them and would not work coming from a smaller account, say so plainly and leave them out of the three.

Make it a habit, not a project

1. 📅 Run it weekly to start. Same morning each week. Weekly fits inside the free Apify credit and gives posts time to finish accumulating views, which makes the scores more stable than a daily run does.
2. 🗂️ Keep every CSV. Save each run dated, for example outliers-2026-08-18.csv. After a month you can ask Claude which patterns keep reappearing, which is a far better signal than any single week.
3. ✂️ Prune the watchlist monthly. If an account has produced no outliers in six weeks, it is not teaching you anything. Swap it out for someone closer to your size who is clearly growing.
4. 🎬 Rebuild the structure, never the script. Take the hook shape and the format, put your own topic, example and opinion inside it. A line-for-line copy reads as a knock-off and almost never performs like the original did.

4 honest limitations

Why I stopped trusting my saved folder

I built this because half the reels I had saved as inspiration turned out not to be outliers at all. They were big accounts having a normal day, or replays doing the heavy lifting, and copying them went precisely nowhere. The detector did not make me more creative. It changed which posts I was willing to spend a week studying, and that turned out to be the decision that actually mattered.