`--voice` repeats, and with none given the script takes the bed by name and every
non-SFX track playing over it. Audio elements are preferred: video counts only when
no audio track is left to be the voice, or every B-roll clip in a composition reads
as somebody talking — which is exactly what happened on the first run against a real
project.
The by-ear tie-break is gone with the single-voice model that needed it: it existed
to choose one voice, and there is nothing to choose now. `--static` is gone too.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Mixing was the one audio surface no skill covered: the effect chain, automation
lanes and the voiceover carve existed with no guidance, so an agent had the
attributes and none of the judgement for using them. The skill teaches the carve
as what it is — a relationship between two tracks, wired like a sidechain, with
the settings on the bed naming the voice — and routes the effect families by the
problem each solves rather than listing parameters.
It ships scripts/carve.mjs because a skill teaching a feature agents cannot
operate is not much of a skill: the analysis needs decoded PCM, and the only other
way in is clicking a Studio panel. Same core functions and same decode rate, so
headless output and panel output are the same three attributes. It finds the voice
and the bed itself — names first, then by ear, measuring how much of each track is
quiet, since a voice stops between phrases and a bed does not — and refuses rather
than guessing when two tracks are too close to call.
References carry the full registry with which parameters can actually be automated
(the four worklet effects expose none, so a lane on one is silently inert) and the
exact JSON of the three attributes.
Counts, catalogues and the core-skill manifest updated in lockstep; the pinning
test in skillsManifest is what caught the two surfaces the maintenance checklist
in CLAUDE.md does not name.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
