broll-catalog-builder（布罗尔目录生成器）

B-roll Catalog Builder

Follow shared release-shell rules in:

• postplus-shared release-shell rules

Use this skill when the user has B-roll assets and needs a reusable asset inventory before planning the edit.

This skill should create the asset mother file for the B-roll side of the workflow.

Default output:

• broll-catalog.json

Optional supporting outputs:

• contact sheets
• frame manifests
• markdown review summaries

Use For

• organizing a B-roll folder into a reusable catalog
• tagging B-roll by semantic meaning instead of filename only
• identifying the best usable ranges within longer B-roll clips
• separating strong proof clips from weak filler clips
• preparing assets for later beat-to-B-roll matching

Trigger Signals

Use this skill when the user asks for things like:

• 先整理一下这些 B-roll 素材
• 给这些素材打标签
• 哪些素材适合做 proof
• 给我一个可复用的 B-roll 素材表
• 先把素材库建起来

Do not use this skill when the user already has a good catalog and needs final beat matching.

Route that to:

• skills/40-creative/b-roll-family/broll-match-engine

Read These References

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

Use these existing local skills when helpful:

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

Core Principle

Do not catalog assets only by what is literally visible.

Also capture:

• what the asset proves
• what edit role it can play
• how long it stays fresh on screen
• whether it is strong enough for primary proof or only support

The same clip may be:

• strong proof
• weak proof but good transition cover
• pure pace reset
• UI demo

depending on how it is used.

Required Inputs

At least one of:

• local folder of B-roll files
• manifest of B-roll files
• list of local file paths

Optional but valuable:

• campaign context
• transcript or script
• target platform
• known proof goals

Workflow

1. Recover the asset set

Build the working file list from:

• folder scan
• manifest
• explicit file list

Preserve local paths.

2. Inspect the assets

For each asset, capture:

• media type
• duration if video
• literal description
• strongest usable ranges
• semantic tags
• support roles
• visual risks

Use lightweight extraction or contact-sheet generation where helpful.

3. Judge edit usefulness

For each asset, decide:

• is it primary proof, supporting proof, or filler
• what claims it can support
• whether it fits 9:16, 16:9, or both
• whether text or UI is readable enough

4. Package the catalog

Build broll-catalog.json using the shared contract.

The catalog should be reusable across multiple edit runs.

Minimal Script

First implementation script:

• scripts/run_build_broll_catalog.mjs

Current first-version behavior:

• scans a local folder recursively
• keeps video and image assets
• uses ffprobe for duration and dimensions
• reads same-basename .md sidecars when present
• infers conservative semantic tags and support roles from filename and sidecar text
• writes broll-catalog.json

Example:

node ${CLAUDE_SKILL_DIR}/scripts/run_build_broll_catalog.mjs \
  --input-dir /abs/path/to/b-roll-source \
  --output <work-folder>/broll-catalog.json

Release-Shell Execution Contract

• keep folder scans, contact-sheet inputs, frame manifests, and intermediate catalog builds under <work-folder>/.postplus/broll-catalog-builder/
• keep only final user-facing broll-catalog.json or review exports outside .postplus/
• start with a bounded first pass on one folder or shortlist before broader cataloging
• if ffprobe or ffmpeg is missing, the user's agent should proactively install the ffmpeg package with the host package manager already present on the machine before continuing
• rerun direct checks such as ffprobe -version and ffmpeg -version after installation
• if installation or verification fails, stop immediately instead of switching to ad hoc shell glue

Output Standard

The catalog is good if a downstream matcher can answer:

• what assets exist
• which claims they support
• which ones are actually strong
• where the useful cut ranges are

If the output is only a file list with vague tags, it is not enough.

First-Version Boundary

Prefer:

• local file scanning
• ffmpeg metadata
• frame extraction when needed
• markdown and JSON packaging

Do not require:

• heavy CV pipelines
• perfect object recognition
• fully automatic clip scoring without review

Review Rule

When confidence is low, say so.

Do not pretend a weak or generic clip is strong proof footage.

It is better to mark an asset as:

• generic filler

than to overstate its usefulness.