The Headliner: turning YouTube title research into a tool a strategist runs in minutes
A video marketing agency was researching YouTube titles for its clients by hand, one search at a time. Four weeks later the research runs as a tool the team owns, with the judgement left to the strategist.
The repetitive task
Humble&Brag produces YouTube content for B2B clients. Before any video is published, someone has to answer the same question: what title will this video need to earn the click? The honest way to answer it is to look at what is already working, in the client's niche and outside it, and find the patterns worth borrowing.
Done by hand that means a strategist typing searches into YouTube, opening channels, eyeballing view counts against channel size, pasting titles into a document and then trying to work out which of them are outliers and why. It takes a long time per video, it depends on who is doing it, and the results are hard to reuse for the next client. The agency had an early prototype that filtered titles with regular expressions, and a competitor product it did not want to pay for. It wanted its own tool, shaped around how its strategists actually think.
What was built
The Headliner takes one seed title and returns four things on a single screen.
-
Topic focused results
Claude reads the seed title, works out the topic, and writes semantic search queries for the YouTube Data API. Each returned video is scored 0 to 100 for topical relevance and anything off topic is dropped. The table shows views, likes, clicks and engagement rates, with Shorts hidden by default.
-
Outlier scoring
Every video's daily views are compared with the other results of the same age (published within a day, a week, a month, and so on) as a z-score, so a three-day-old video is never judged against a three-year-old one, and a fresh video from a small channel can be recognised as a breakout. Outliers are graded mild or strong, and the whole table sorts by outlier score by default. The agency can define that a title on a channel of any size "punched above its weight", which is the signal they care about.
-
Cross niche pattern matches
The part that took the most iteration. Instead of keyword matching, the tool first breaks the seed title into its structural formula (imperative plus time span plus outcome, question plus number, and so on), then searches other niches for strong positive outliers that follow the same formula. Niche filter chips let the strategist exclude categories such as gaming that dominate the results but do not translate to B2B.
-
AI generated titles and the Headliner
Ten title suggestions, weighted towards the patterns in the breakout titles rather than generic advice, with the single strongest marked as the Headliner. An Insights section explains the patterns, hooks and keywords the suggestions are built on, so the strategist can disagree with the reasoning rather than just the output.
What did not work the first time
The first version of the AI suggestions was, in the client's words, "very ChatGPT": fluent, plausible and far from what actually performs. Two changes fixed it. The outliers were put into the prompt as a dedicated breakout section with a short performance summary, so the model was reasoning from titles that had demonstrably worked rather than from its own idea of a good title. And the model was switched from OpenAI to Claude, which used the agency's custom instructions more faithfully and fell back on generic patterns less.
The cross niche problem
Keyword similarity finds videos about the same thing. The agency wanted videos built the same way.
The first cross niche table matched on shared words and returned noise. The rewrite extracts the syntax defining words of the seed title ("Why", "Every", "Should", "Give me 10 minutes and"), checks their order and position, and rewards results that follow the same structure even from a completely different topic. With a stricter threshold the table sometimes shows fewer than ten rows. The rows it does show are the ones worth learning from, which was the point.
The other practical problem was YouTube API quota. Semantic search means several queries per seed title, and the free quota ran out during testing. Two fixes: results are cached for seven days so the same search does not burn quota twice, and a second API key under a separate Google Cloud project doubles the available quota. Both are documented so the agency can repeat the pattern if usage grows.
Human in the loop, by design
The tool does the gathering, scoring and first draft. The strategist does the choosing. Nothing is published from the Headliner; it produces a shortlist with reasons, and the person who knows the client decides. Access is by invite code with admin roles, so the agency controls who uses it and can see who registered, which matters because every search spends API budget.
Everything runs in the client's accounts: their Supabase project, their Netlify site, their Google Cloud project, their model API key with a spending cap and auto recharge switched off. If the engagement ended tomorrow, the tool would keep working and the agency would keep the bill, the data and the code.
What this unlocks
The agency went from hand research per video to a repeatable process it can run for any client in a few minutes, with the same scoring logic every time. Within the first fortnight of testing the co-founder reported that the tool had already started saving him time, and two months in that it had been incredibly helpful for ideating for the agency's own clients and he already considered it a big success.
The next tool is already scoped.
A task your team repeats by hand?
Research, reporting, triage, classification: if a process has a clear input, a clear output and a person who should still make the final call, it is a candidate. Start with an implementation audit to find the one or two tasks worth building, or describe the task if you already know it.