Scene 2 — the stadium | Academy

Every scene so far has kept the hero in his casual look — pink shirt, white tee, jeans, locked back when the character sheet was built. That's fine for a kitchen; it's wrong for a running track.

Build a second, matching outfit

The hero's casual look — pink shirt, white tee, jeans. Locked for the kitchen, wrong for a stadium.

Instead of hand-writing a new prompt, the casual sheet and a sneaker reference go straight to Claude with one request: edit this character into a blue athletic outfit wearing these sneakers. That prompt goes into GPT Image 2 along with the same two references, and a few clean, consistent options come back. One gets picked and locked in as the dry version of the hero.

The locked dry hero — navy athletic tee and shorts, the same sneakers, front and back. This is the reference used for the warm-up and the run.

Two sheets, not one prompt fighting itself

The scene needs him dry for the warm-up and the run, then drenched in sweat by the finish — and that's exactly the kind of mid-scene change one reference can't hold. Ask the model to add sweat with words on top of the dry sheet and it starts improvising: the face drips, details go wrong. So a second sheet gets built instead — same guy, same outfit, described as just off a run and soaked through, sweat stains on the shirt — and locked as its own reference.

The locked wet hero — same navy outfit, now sweat-darkened down the chest. Built as its own reference instead of editing sweat onto the dry sheet.

If a character has to change mid-scene, don't fight one reference with more words — build a second reference instead. Images are cheap; video generations aren't.

Claude is then told exactly when to use which: the dry hero for the warm-up and the run, the wet hero only for the final cut where he's soaked.

Override the scene's own lighting

The stadium location, locked earlier in the build. This scene overrides the shot list's default light instead of inheriting it.

The style prefix set earlier — soft, even daylight, right for the kitchen — is wrong for a stadium at midday. So it gets overridden for this scene only: a hard, high sun behind the camera, a deep blue sky, hard-edged shadows on the track, saturated color. The prefix is just a default; when one scene needs its own look, override just that scene instead of touching the whole shot list.

Fix the cut, not just the shot

Two more notes go back into the run before it's usable: drop the warm-up stretching so the scene is dynamic from the first frame, and match the opening hand-tap to the exact tap that closed the kitchen scene — same hand, same motion — so the cut between scenes reads as one continuous action instead of two clips glued together.

Add coverage, then one shot that's just the product

To make the run worth watching, a second version adds broadcast-style wide shots from the stands, tight tracking on his legs driving, and a worm's-eye view from track level as he spins straight over the lens.

The scene is still missing the one thing this whole ad is actually selling: a shot where the headphones, not the hero, are the subject. A third version keeps the same opening and closing tap but swaps the middle for a snorricam-style body rig — a camera locked to his side, bolted to the right ear cup, the headphones dead center and rock solid while the whole track blurs past behind them.

That kind of shot normally means an operator, a rig, a location day, and a few thousand dollars before anyone hits record. Here it's one written prompt.

Three prompts, six generated videos — and the stadium scene is cut from pieces of all of them: the dynamic wide and the worm's-eye from one clip, the locked product shot from another, the finish from a couple more.