Someone presses one button
The report doesn't run on a timer. On Wednesday a person presses a button on a private dashboard. That opens a session with Claude, Anthropic's AI, and hands it a written checklist.
The checklist is the same every week, so every report is made the same way. When the run ends, the person gets a note saying what was found and anything that couldn't be checked.
Behind the scenes: the checklist is a Claude skill called trend-report. It holds the rules too, like "only state facts you checked this week".
Every trend has to earn its number
Claude starts with the latest round-ups from New Engen, SocialPilot, Buffer, Postdom and others to see what people are talking about. A mention isn't enough to get on the list.
It then opens the trend's own TikTok sound page or Instagram audio page and reads the count from there. If only a round-up has the number, the page names that round-up. If nothing can be checked, it's left out.
Behind the scenes: each trend becomes one entry in data/trends.json with what it is, where it came from, its sound, a brand idea, what to watch out for, example posts and sources.
The list puts usable trends first
Trends a brand can run any week go at the top, ranked by how many videos use them. Trends tied to a date go underneath, ranked the same way.
That's why a huge Christmas sound can sit below a smaller dance. Once its date passes, it's no use to anyone.
Behind the scenes: a trend keeps its exact name from week to week. That's how past reports can show when it first appeared and how much it has grown.
Each trend gets a status
Rising, Peaking, Steady or Past peak. It tells a brand whether there's still time to join in.
This part is a judgement. Claude compares the count with earlier dated counts. Where there's only one count, it looks at how recent the biggest posts are and whether new ones are still landing.
Behind the scenes: the reasoning behind each status is saved with the trend, so a call can be questioned later.
Pictures come from real posts
Each trend gets one picture: the cover frame of a real example post, linked back to that post.
A script opens the platforms' public pages with no login and saves the frame. If no post can be captured, an AI illustration fills the gap and the page labels it as one.
Behind the scenes: tools/shots.cjs collects the frames and tools/ai_images.py fills any gaps.
Slang is measured by dictionary lookups
Nothing free counts words in TikTok comments, so the tracker uses a stand-in: how many people look a word up on Wiktionary, from Wikimedia's public figures. Each word's last 14 days are compared with the 14 before and with its best stretch in 26 weeks.
- Rising
- Up 20% or more on the 14 days before
- Peaking
- Within 15% of its best, and at least 25% above its usual level
- Past peak
- Down 20% or more, or at 60% or less of its best
- Steady
- Anything else, or under 40 lookups in 14 days
Behind the scenes: unlike trend status, this is a fixed rule in tools/words.py, with no judgement in it. Phrases have no dictionary page, so their status is set by hand.
One file becomes the site and the slides
Claude writes the week's headlines into one file. A script builds this site from it and freezes a copy in past reports, so old weeks never change. Pushing to GitHub Pages puts it live about a minute later.
The same file fills a fresh copy of the slide template, pictures included. Every slide is then exported and looked at for spilled text, leftover placeholders and the wrong picture.
Behind the scenes: tools/build.py makes the pages and tools/deck.py writes the changes for Google's Slides API. It stops if a line is too long for its box, because Slides won't shrink text to fit.