Direction Desk

Direction, not prompting. The first take is never the final one.

0:00.0

The Marlow spot plays its real rendered voices, and the block lengths follow the actual clip durations. Edited lines and the second script play silently as captions — the current line is highlighted while the timeline runs. Overlap and interruption play as a tight cut-in.

Concept and sources

Modern speech models read inline performance tags — the vocabulary for delivery exists as syntax, but there is little interface for directing with it. Direction Desk treats those directions as objects a director places, auditions and approves, instead of text to be typed. Its take board follows the alternatives-first pattern: aligned variants with explicit differences, shown to help editors judge generative media faster than one-result-at-a-time review.

The Marlow spot and its take board play real rendered audio, embedded in this page. Lines the studio has not rendered — edited lines and the second example script — play silently as captions.

Sources: VideoDiff, Adobe Research, Conference on Human Factors in Computing Systems, 2025.
RockBloom. Labs Part of the Anthropic Startup Program