broll-match-engine（布罗尔匹配引擎）

B-roll Match Engine

Follow shared release-shell rules in:

• postplus-shared release-shell rules

Use this skill when the user wants to map a talking-head edit to candidate B-roll coverage.

This skill should produce a structured match plan, not a vague creative summary.

Default output:

• broll-plan.json

Optional supporting outputs:

• markdown match summary
• beat review table
• candidate ranking table

Use For

• deciding which spoken beats need B-roll
• matching a beat to the best available proof clip
• deciding when to stay on A-roll instead of cutting away
• identifying missing proof moments
• attaching keyword-emphasis or motion-hint metadata to the match result

Trigger Signals

Use this skill when the user asks for things like:

• 这些口播句子该配哪些 B-roll
• 帮我匹配素材
• 哪些地方应该切屏幕或证明画面
• 这条 talking head 怎么上 B-roll 更合理
• 给我一个 B-roll plan

Do not use this skill when the user still lacks a usable B-roll inventory.

Route that first to:

• skills/40-creative/b-roll-family/broll-catalog-builder

Read These References

• family architecture: ../references/skill-architecture.md
• shared contracts: ../references/contracts.md

Use these existing local skills when helpful:

• a dedicated visual analysis workflow
• skills/40-creative/editing-decision-engine

Core Principle

Do not match by keyword overlap only.

A good match depends on:

• beat meaning
• proof need
• visual support type
• pacing need
• B-roll strength

The best asset is not always the most literally similar one.

Sometimes the right choice is:

• a UI proof clip
• a workflow bridge clip
• a pace reset clip
• or no B-roll at all

Preferred Inputs

Required:

• normalized-transcript.json
• broll-catalog.json

Strongly preferred:

• chunked-<mode>.json

Optional but high-value:

• A-roll semantic analysis
• reference edit pattern
• target platform
• intended output length

Workflow

1. Build beat units

Prefer subtitle chunks or edit beats over raw transcript segments.

For each beat, capture:

• spoken text
• timing
• beat role
• visual need

2. Evaluate B-roll need

Not every beat deserves a cutaway.

For each beat, decide:

• stay on face
• use supporting B-roll
• use primary proof B-roll
• use light visual reset only

3. Rank candidate assets

For each beat that needs B-roll, rank candidates based on:

• semantic fit
• support-role fit
• proof strength
• pacing fit
• visual clarity
• usable range quality

Every candidate should include a plain-language reason.

4. Package the match result

Build broll-plan.json with:

• beat metadata
• B-roll need decision
• candidate list
• fallback
• keyword overlay suggestions
• motion hints

Minimal Script

First implementation script:

• scripts/run_match_broll_plan.mjs

Current first-version behavior:

• reads chunked-<mode>.json
• reads broll-catalog.json
• treats subtitle chunks as beat units
• infers beatRole and visualNeed from spoken text
• ranks B-roll candidates using conservative heuristics over:

- spoken keywords - catalog semanticTags - catalog supportRoles

• writes broll-plan.json

Example:

node ${CLAUDE_SKILL_DIR}/scripts/run_match_broll_plan.mjs \
  --chunks /abs/path/chunked-basic.json \
  --catalog /abs/path/broll-catalog.json \
  --output <work-folder>/broll-plan.json

Release-Shell Execution Contract

• keep beat chunks, candidate rankings, and intermediate match plans under <work-folder>/.postplus/broll-match-engine/
• keep only final user-facing broll-plan.json or review exports outside .postplus/
• start with a bounded first pass on one transcript plus one catalog before broader edit planning
• if required upstream inputs are missing or invalid, stop immediately instead of switching to ad hoc shell glue

Output Standard

A strong output should let a human editor answer:

• where to cut away
• what to cut to
• how long to stay there
• what the fallback is if the preferred clip is weak

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

First-Version Boundary

Prefer:

• explainable candidate ranking
• beat-level JSON
• optional markdown review

Do not require:

• fully automatic timeline assembly
• NLE export
• guaranteed perfect semantic matching
• motion graphics rendering

Review Rule

Always distinguish:

• essential proof
• useful support
• optional filler

If the inventory does not contain strong proof footage, the plan should say so directly.