Studio

The source was a silent screen recording just under six minutes long. Contact sheets made the sequence visible. The first agent edit selected eleven spans, compressed navigation and waiting, and held consequential terminal states long enough to read.

I shortened that first cut. My preferred sequence then became the picture reference: ten source spans, each with its own speed and output interval. The reconstructed picture runs for about 1 minute 59.5 seconds.

The narration arrived as separate takes. The agent transcribed and timed them, selected seven source spans, and assembled them by meaning and performance. A firmer replacement sentence could be inserted without rerecording the rest.

The resulting map connects each piece of narration to the event it explains. Pauses allow the screen to resolve and preserve sentence boundaries. Cleanup used filtering, equalisation, gating and loudness normalisation, with the selected performance guiding the edit.

Early crops narrowed the terminal and lost Shasta at the edge of the frame. The later edit returned to the original recording, recovered the full width and preserved Shasta alongside the demonstration.

A later review found an unwanted low-frequency sound near the sign-off. Waveform and spectral views helped isolate it from the surrounding speech. The repair replaced the 110 milliseconds from 113.990 to 114.100 seconds, with short fades around the join.

The spoken ending remained intact. The production record's sample comparison confined the difference to the repair and its fades, while the encoded picture payload stayed unchanged.

An AI-narrated alternative reused the approved visual timing. Seven generated sections were fitted to semantic slots, with pronunciation directions kept separate from the canonical spelling in the captions.

One dense section sounded processed when stretched. Its replacement kept the speech at natural speed and distributed 4.19 seconds of silence across existing sentence boundaries. That small editorial change preserved the timing without stretching the words.

Transcripts and timed captions were reconciled against the finished performances. The retained human, repaired-human and AI editions share the same encoded picture payload.

Each correction reopened a particular part of the edit: a sentence, a crop, a pause or a fraction of a second of audio. The accepted picture, wording and performance stayed available as explicit references for the next operation.

The continuing Studio work asks how to make that process reusable: accept fragments, show an editable understanding of their relationships, and preserve the parts already accepted while refining the rest.