media-router（媒体路由器）

Media Router

Follow shared release-shell rules in:

• postplus-shared release-shell rules

Use this skill when the user wants to do something with:

• audio
• video
• transcripts
• subtitles
• A-roll / B-roll edit prep

This skill is a routing layer.

It should decide:

• what the actual task is
• what output shape is required
• whether the job is rough, subtitle-ready, or edit-ready
• whether cost or quality should dominate
• which downstream skill or skill chain should execute

This skill should not pretend every request is "just transcription".

Use For

• deciding whether to call audio STT or video STT
• deciding when timestamps are required
• deciding when subtitle packaging is enough
• deciding when semantic video understanding must be added
• deciding when editing-decision-engine should run
• deciding whether a batch workflow should be used

Read These References

• references/brief-schema.md
• references/routing-modes.md

Core Rule

Route by job intent, not by input file type alone.

A video input may still be:

• plain transcription
• subtitle generation
• transcript plus semantic understanding
• edit-beat planning

The route should be chosen from five dimensions:

• input type
• output goal
• required precision
• scale
• cost mode

Default Routing Logic

audio -> transcript

Use skills/40-creative/audio-transcription.

video -> transcript/subtitles

Use skills/40-creative/video-transcription.

transcript with timestamps -> subtitle files

Use skills/40-creative/subtitle-packager.

transcript + assets -> cut logic / B-roll plan

Use skills/40-creative/editing-decision-engine.

video -> edit-ready plan

Chain:

1. video-transcription
2. subtitle-packager if subtitle files are needed
3. frame-extraction if visual-proof snapshots are needed
4. editing-decision-engine

Quality Thresholds

rough

Good enough for:

• search
• quick review
• rough logging
• candidate beat spotting

Prefer:

• cheaper models
• fewer enrichments

subtitle-ready

Good enough for:

• human-readable captions
• SRT / VTT generation
• reliable sentence or word timestamps

Prefer:

• Whisper with timestamps

edit-ready

Good enough for:

• A-roll/B-roll decision making
• cutaway placement
• beat-level planning

Prefer:

• Whisper with timestamps
• semantic video understanding when proof footage matters
• editing-decision-engine downstream

User-Facing Explanation Rule

Do not expose internal route labels unless the user is explicitly designing the system.

Normal user-facing language should say:

• what will be extracted first
• what will be aligned second
• what editing decisions become possible afterward

Default Output

Return:

• chosen route
• why it fits the task
• primary skill
• supporting skills
• model plan
• expected output artifacts
• whether timestamps are mandatory