Video Rough Cut

Use this skill for fast rough-cut cleanup of a single talking-head or voiceover video using the local B-Roll Studio rough-cut pipeline. It uploads one raw vi...

Security Scan

The skill is internally consistent with a local rough-cut workflow and contains a simple client script that posts a local video to a localhost API; there are no unexplained credential requests or remote downloads.

Purpose & Capability

Name/description, SKILL.md, api.md, pipeline.md, and the Python client all describe the same local rough-cut use case: submit a single talking-head video to a local B-Roll Studio API, poll for completion, and download the output. The required capabilities (reading one local video and talking to a local API) match the stated purpose.

Instruction Scope

SKILL.md instructs the agent to run bundled resources at paths like scripts/run_rough_cut.py and references/references/api.md, but the manifest contains run_rough_cut.py, api.md, and pipeline.md at the repository root (no 'scripts' or 'references' directories). That mismatch will likely cause runtime failures or confusion. Otherwise the instructions are scoped to the stated task and do not request unrelated files, environment variables, or external endpoints.

Install Mechanism

There is no install spec — this is instruction-plus-client-script only. No packages are pulled from remote hosts and no archives are extracted. The client uses Python stdlib (urllib), so no additional install privileges are required.

Credentials

The skill requests no environment variables, no credentials, and no config paths. The script's network activity is limited to a configurable base_url (defaulting to http://localhost/api/v1), which is consistent with the described local API integration.

Persistence & Privilege

The skill is not 'always: true', does not claim to modify other skills or system settings, and contains no persistent install steps. It runs a one-off client that reads a local video and communicates with the local API.

Assessment

This skill appears to do what it says: it posts a local video to a local B-Roll Studio API, polls the job, and downloads the result. Before installing/using it: (1) fix the path mismatches in SKILL.md (it refers to scripts/run_rough_cut.py and references/*.md but the files are at the root) so the agent can find the bundled script and docs; (2) only run this if you trust the local B-Roll Studio service — the client will upload any file you point it at to the configured base_url (the default is localhost, but base_url can be overridden); (3) confirm you are comfortable the agent can read the input video file (it requires local file access); and (4) if you plan to run this in automated/autonomous mode, ensure it cannot be redirected to a remote base_url you don't control. No credentials are requested by the skill.

Use this skill when the user has one raw talking-head /口播视频 and wants a cleaned first-pass edit, not a full B-roll edit.

Default output:

• one cleaned draft video
• optional cut-decision review data

Inputs

Required:

• one local video file

Optional:

• platform preset (douyin, etc.)
• whether to remove pauses
• whether to remove breaths
• whether to trim head/tail clutter
• whether to stabilize, auto-center, adjust brightness, apply beauty

Preferred execution path

Prefer the local B-Roll Studio rough-cut API if the app is running.

Base URL:

• http://localhost/api/v1

Use the bundled script:

• scripts/run_rough_cut.py

Read references/api.md when you need endpoint details. Read references/pipeline.md when you need to explain what the current rough-cut logic actually does.

Workflow

1. Validate prerequisites

Confirm:

• the input video exists
• local B-Roll Studio is running if you plan to use the API

2. Submit a rough-cut job

Use the bundled script:

python3 <skill-path>/scripts/run_rough_cut.py \
  --video "/absolute/path/to/video.mp4" \
  --base-url "http://localhost/api/v1" \
  --wait \
  --download

3. Default processing assumptions

Unless the user says otherwise:

• remove pauses: on
• remove breaths: on
• trim head/tail clutter: on
• stabilize: on
• auto-center: on
• brightness adjustment: on
• beauty mode: light
• denoise audio: off

4. Interpret the output correctly

This pipeline is for mechanical cleanup, not final editorial polish.

It is good for:

• removing empty pauses
• trimming 321走 / 试音 / 收镜头动作
• producing a cleaner draft for review

It is not the right tool for:

• detailed B-roll placement
• manual storytelling edits
• precision subtitle design

5. If the user reports quality issues

Typical corrective moves:

• 正文被裁掉: reduce head/tail trimming confidence or review cut decisions
• 尾部收镜头没裁掉: inspect cut decisions and rerun with trim enabled
• 声音发闷: keep denoise off
• 画面太亮: keep brightness adjustment on; the current pipeline clamps over-bright footage instead of only brightening

Quality bar

The result should feel like a clean draft editor would hand off for review:

• no partial trim of 321走
• no obvious trailing mirror-check / 收镜头动作
• pauses removed without harming正文
• brightness stays natural
• no heavy-handed denoise by default

Fallback

If the local API is not available:

• explain that this skill depends on the local B-Roll Studio rough-cut service
• provide the API health check command
• do not invent a separate rough-cut algorithm unless the user explicitly asks you to rebuild one

Bundled resources

• scripts/run_rough_cut.py Submit, poll, inspect, and download local rough-cut jobs.
• references/api.md Rough-cut API endpoints and request fields.
• references/pipeline.md Current rough-cut architecture distilled from the project.