Imported from lukemun/opensequence (.agents/skills/music-audition/SKILL.md). Install upstream with npx skills add lukemun/opensequence --skill music-audition. Copyright stays with the author.

Music audition

Read ../music-direction/SKILL.md and ../timeline-auditions/SKILL.md before editor work, including their proposal guidance. Search existing music before fetching new tracks.

Find passages

Read the configured Open Sequence music catalog (TIMELINE_MUSIC_CATALOG_DIR, otherwise music-catalog/catalog.json). Also search the sibling Social Content library: CATALOG.md, catalog.csv, and shoots/*/annotations/annotations.json. Those annotations retain full-source hashes, review status, cue windows, and provisional tags that the editor's smaller schema cannot express. Resolve paths relative to the containing catalog, never from a filename alone. If that library is unavailable, report the missing search scope and use the available catalog.

Match the actual request against cue-level mood tags, energy trajectory, turn/peak offsets, available duration, and narration needs. A matching track title alone is insufficient. For provisional tags, state the uncertainty and listen/review before treating mood, phrase boundaries, vocals, or tempo as confirmed. Do not automatically rerun extraction when only the search brief changes.

If no scene or plan exists, browse a few relevant passages with source timings; do not invent a production brief or select a winner. Catalog preparation belongs in Social Content's music-cataloging workflow and can finish without a timeline.

Grow the catalog from a real edit

When the available catalog has no credible fit, use the current edit as a narrow acquisition brief rather than starting a broad music search. Add only exact recordings that can be reused: verify the composition and the recording separately, preserve the performer or recording authority, source page, license, file hash, full duration, and any attribution or delivery obligation. A public-domain composition does not make an arbitrary recording public domain.

Create one cue window for the passage actually auditioned, including its source-in/out, intended role, measured comparison gain or loudness note, and the picture/dialogue boundary it was designed around. Keep mood, tempo, phrase, and speech-compatibility judgments explicitly provisional until a human hears the cue both solo and against the edit. The first project's choice is evidence for that use case, not universal proof; broaden good_for or promote additional cue windows only after later edits demonstrate them.

Prefer a low-obligation recording when two options are editorially equivalent. Never hide attribution, share-alike, platform-only, territory, or reference-only constraints merely to make an option easier to ship.

Make the comparison useful

Use a small shortlist (usually 3–5 materially different cues). Label each option with track/version, source-in/out, duration, proposed emotional role, and why it matches. Preserve reference-only status; do not relabel it licensed or invent a platform sound mapping.

For a scene audition, compare the same picture and dialogue window with each music option. Align the relevant musical turn or peak to the same scene beat; record the cue's source offset and resulting timeline offset. Match perceived playback loudness with a documented per-option gain, retaining each cue's internal dynamics. Avoid judging fit from a louder option. Keep narration intelligible and allow hearing each cue alone as well as in context.

Use the current audition_create schema for supported net-new candidates, or proposal options for an existing working passage. Read capabilities before assembling; the historical scripts/build-music-audition.ts uses direct timeline writes and is not an authorized shortcut. Keep music as editable audio. If the current schema cannot represent a source status or comparison, preserve a standalone playable shortlist and report the editor limitation rather than coercing fields.

For a standalone audition, the Social Content cue browser provides full-track context and separate excerpts without requiring a project. Verify media loads, playback advances, windows stop at their ends, and source mappings are correct before calling a comparison playable. A script or waveform alone is not a listening review.

Record the choice

Leave all options unselected until the user chooses. Record the chosen track/version/cue IDs, original source hash and interval, comparison gain, alignment anchor, and the user's reason when supplied. Keep this decision with that project's brief; a choice for one scene is not a global taste preference. Carry the selected source window into a reviewable placement proposal. Selection does not itself authorize publishing or deleting alternatives.