05 · Method
Three finished cuts · 1080×1920 · one shipped
The flagship ad went through several builds before it shipped. Three of them are worth showing, because the difference between them is not production quality — it is who is on screen.
The first strips the presenter out entirely and runs narrator-led. It works, which established that the structure was sound and the problem was casting, not construction. The second is a build-vs-buy control: an original character, designed and prompt-directed in-house, then rendered end to end by a paid vendor. The third recasts deliberately for the audience — an applicant at a kitchen table with a study notebook, not a presenter and not an influencer.
The third shipped. Narrator-led became the standing fallback for products with no fitting persona, and the vendor route was declined on a specific finding: it reaches comparable persona quality but cannot carry real in-app footage, which is the proof beat every one of these ads is built on.
Tone was the expensive variable. The influencer read is technically clean and tonally wrong for this audience — confident and polished, aimed at people who are anxious and skeptical. It reads as a sales pitch and undercuts the product.
The faceless cut solves that by removing the problem rather than fixing it. It is a legitimate deliverable and the right answer when no persona fits, but it gives up the strongest thing short-form has: someone who looks like the viewer.
The third recasts to that. Same three-beat structure, same runtime, same pipeline — the only change is who opens the ad, and it is the version that shipped.
Structured A/B iteration · QC criteria definition · persona design and re-casting · vendor build-vs-buy evaluation · silence detection and trimming.