How Claude built a singing roof

There is protein in coffee and protein in snacks. Why not in a roof? That absurd idea is the starting point of the 43-second video I published on the Talya Roofing channel. A house sings, a package lands on its roof, and it grows muscles while “150 grams” appears on screen. Then the show stops and the video answers its own question: does your roof actually need protein? “No.”

It is easy to say, “Claude made this.” But that sentence mixes up what already existed with what was made for the video. The song was already there. Claude's contribution lies in the details of building moving images around it: which beat brings in a shingle, when the screen clears, and how the singing house's mouth works. I want to look closely at those three examples. They explain the work better than a list of tools because they show how specific decisions shape the picture and can be revised later.

The YouTube title is Everything Has Protein Now. Even Your Roof? You can watch it below. The player connects to YouTube only after you choose to load it.

PROTEIN ROOF▶LOAD VIDEO PLAYER YOUTUBE / SOUND ON YOUR TERMS

YouTube loads after you choose to open the player.

Watch on YouTube ↗

The song existed; the scene timing did not

The starting file was an audio recording called “150 grams of Protein,” about 43 seconds long. I cannot credit Claude with composing the song or generating the vocal in this production. The existing recording was analyzed for words, beats and changes in energy to build the animation's timeline.

That requires two kinds of information. First, when each word is sung. A raw transcript produced with faster-whisper running locally was kept separately from the corrected lyrics. The brand name being misheard and then corrected to Talya Roofing is a small but decisive example. Showing a caption at exactly the right moment does not make the wrong word correct.

Second, the movement of the music. Librosa was used to analyze tempo and beats, with an estimated tempo of about 129 BPM. Building the video takes more than knowing whether the song is fast or slow, though. It needs the exact location of the beat that triggers movement on the roof. Alongside playing behind the picture, the song becomes a measure for when the scenes change.

A drum hit becomes a row of shingles

The first example is the house's transformation. It stands under its own small raincloud with shingles missing from its roof. A package arrives, the old covering scatters, and new shingles fly into place. In the source code, those arrivals are tied to snare hits. As the protein counter climbs, the roof fills in with the music.

That gives the song's exaggerated promise of strength something visible to point to. The house does not simply appear repaired in the next shot; we watch it change piece by piece. Seeing the same character soaked by rain and then flexing its muscles also lets the story move without a long explanation.

At around 23 seconds, the chorus makes everything bigger: a sunburst background, large lettering and the flexing house. A few seconds later, the opposite happens. A parody supplement label gives way to a plain question card, followed by the single word “No.” on black.

That pause is the second example. If every scene kept the same intensity, the answer might read as just another piece of text. Clearing the screen gives it its own weight. The next line, “But it might need Talya Roofing,” connects the protein joke to the roofing business. Timing does more than keep the words and pictures together here: it builds the claim, then makes room to puncture it.

Changing the mouth also changed the framing

The third example is the difference between two production versions. In the first, the front door moved like a mouth. In the published version, the door stays in place and a separate cartoon mouth sits between it and the eyes. That change shows Claude's contribution more clearly than a single finished frame: the character's construction could be revised, and the same scenes could be rendered again with the revision.

The house and surrounding objects were drawn in Python. Windows act as eyes, blinds as eyelids, and awnings as eyebrows. Pycairo draws the shapes, while the character's position, scale and expression are calculated over time. The existing Talya logo is placed within those drawings. The mouth therefore becomes a moving part of the character rather than a detail pasted over the finished picture.

The new mouth uses word timings and approximate syllables. It closes for some sounds and opens or rounds for others; vocal energy also affects how far it opens. This is not a system that analyzes speech perfectly. It maps spelling to mouth shapes using rules built for this character. Between sung words, a smile or a sad expression takes over according to the scene.

Making room for the mouth made the house taller. In some scenes the character was then lowered or reduced in size, keeping the roof clear of large lettering. That chain of adjustments matters: changing the mouth meant checking the whole composition. It shows the practical value of scenes built in code. The character can change, but its relationship to everything around it still needs another look.

Returning to a finished video

There are three opening versions. I published the one beginning with “Everything has protein now.” The others change the opening text while keeping the main video the same. There are no comparative viewing results here; having three files does not establish which performs better. The concrete possibility is changing the opening without redesigning the whole video.

FFmpeg combines the final frames with the source audio. The master is 1080 × 1920 at 30 frames per second. But the video file is not the only thing left: word timings, scene drawings and character movements remain in separate sources. As the mouth revision demonstrates, a detail that looks finished can be reopened and produced again.

That is why, in describing this video, I focus less on how many tools Claude used and more on how a detail could be changed. The existing song kept its duration while the house's face changed, and the framing was adjusted along with it. Having that source alongside the output shows where the next revision can begin. The first result does not have to be the final word.

The roof does not need protein. But the song's absurd idea needs a character and an answer delivered at the right moment. Claude built the character and timed how the song's answer appears on screen, leaving a structure that could stage the same joke again when the character's face changed. That is what I mean by “made the video”: giving an existing song a picture, and keeping that picture open to revision.