editing-decision-engine（编辑决策引擎）

Editing Decision Engine

Follow shared release-shell rules in:

• postplus-shared release-shell rules

Use this skill when the user wants a strong short-form edit plan, not just a script rewrite or a generic video breakdown.

This skill is for post-edit decision making across:

• A-roll performance
• B-roll proof footage
• script meaning
• beat timing
• reference-video editing patterns
• final edit packaging

Do not treat this as a video renderer.

First-version goal:

• output a high-quality edit decision package
• not fully automate the final NLE timeline

Use For

• planning how to cut a talking-head short
• deciding where B-roll should appear
• mapping script beats to visual proof
• deciding when to stay on A-roll vs cut away
• translating a reference video into reusable edit patterns
• packaging a cut plan for Premiere, Final Cut, CapCut, or a human editor

Trigger Signals

Use this skill when the user asks for things like:

• 这条怎么剪
• 根据 A-roll 和 B-roll 设计剪法
• 参考某条视频的剪辑逻辑
• 按脚本和素材做 post-edit plan
• 逐句决定哪里贴屏幕、哪里留脸
• 给我一个时间轴级别的剪辑方案

Do not use this skill when the user only needs:

• raw video analysis without edit decisions
• frame extraction without timeline reasoning
• render generation from image and audio

Route those to:

• a dedicated visual analysis workflow
• skills/40-creative/frame-extraction
• skills/40-creative/video-batch-runner

Core Principle

Do not describe footage only by what appears on screen.

For edit planning, the important question is:

• what does this shot prove
• what beat does it support
• how much attention should it take
• whether it should lead, support, bridge, or punctuate

The same B-roll clip can be:

• primary proof
• support proof
• transition cover
• pace reset
• visual filler

depending on the spoken beat around it.

Read These References

• workflow and decision rules: references/workflow.md
• core objects and output shapes: references/schemas.md

Required Inputs

The skill works best when at least these exist:

• script text or spoken transcript
• local A-roll file or a reliable description of the A-roll performance
• local B-roll files or a usable B-roll shot inventory
• intended output length or target platform

Optional but high-value inputs:

• one or more reference videos
• subtitle draft or transcript with timestamps
• previous edit notes
• campaign or persona context

If assets are missing, do not pretend the plan is precise.

Instead:

1. state which decisions are grounded
2. state which decisions are provisional
3. produce the strongest plan possible from available evidence

Workflow

1. Lock the edit thesis

Before building a timeline, identify:

• what the video is trying to prove
• whether it is tutorial-led, viewpoint-led, comparison-led, or proof-led
• what visual contrast drives the piece

Examples:

• old fragmented workflow vs in-context workflow
• tool overload vs cleaner setup
• annoying to start vs easy to start

2. Break the spoken content into beats

Do not plan edits sentence by sentence only.

Break into edit beats based on:

• meaning shift
• emotional shift
• proof need
• pace change

Each beat should capture:

• the spoken line or paraphrase
• timing if known
• beat role
• proof requirement
• visual demand

3. Understand assets semantically

For A-roll:

• delivery energy
• pauses
• emphasis words
• facial or body moments worth preserving

For B-roll:

• what it literally shows
• what it proves
• where it is strongest
• how long it can stay on screen before feeling repetitive

For references:

• what structural pattern is reusable
• what is surface style only

4. Decide the cut logic

For each beat, decide:

• stay on A-roll or cut away
• which B-roll asset to use
• whether the B-roll is primary or supporting proof
• overlay length
• subtitle density
• whether to use punch-in, hold, montage, J-cut, or L-cut

5. Package the output

Default output should include:

• edit thesis
• beat map
• B-roll assignment table
• time-ordered edit decisions
• risks and missing assets

Release-Shell Execution Contract

• keep edit theses, beat maps, asset notes, and intermediate decision packages under <work-folder>/.postplus/editing-decision-engine/
• keep only final user-facing edit plans outside .postplus/
• start with a bounded first pass on one sequence or one short video before broader edit planning
• if required inputs such as transcript, A-roll context, or B-roll inventory are missing, stop immediately instead of switching to ad hoc shell glue

First-Version Boundary

Keep the first version pragmatic.

Prefer:

• markdown edit plans
• JSON beat maps
• B-roll proof tables
• simple timeline-ready CSV if useful

Do not require in v1:

• direct XML generation for NLEs
• automatic cut rendering
• perfect shot detection
• automatic motion-design systems

Output Standard

A strong result from this skill should let a human editor start cutting immediately.

Minimum bar:

• they can tell what the first 3 seconds should do
• they know where the proof moments land
• they know which B-roll is essential vs optional
• they know where the edit should breathe instead of over-cutting

If the output cannot guide a real editor, it is not specific enough.