Short Video Content Audit

The Problem It Solves

Creators of short-form video content face significant compliance hurdles before publishing. Platform policies and regulations impose strict rules on text, visuals, and audio, making manual auditing time-consuming and error-prone, especially for multimodal content where voiceover, on-screen elements, and background sounds intersect. A core pain point is the lack of systematic tools to automatically detect potential violations—such as sensitive phrasing, inappropriate imagery, or audio elements—preemptively, thus avoiding content removal or account penalties.

How the Skill Works

This skill automates compliance auditing with a focus on multimodal content analysis, covering text, visuals, and audio. Its workflow includes:

• Input Collection: Users must provide at least one input: oral/subtitle text (recommended ≥20 characters), a local video file (≤100MB, supporting formats like mp4/mov), or an existing video_id. Optional fields like platform (e.g., Douyin, Video Accounts) and industry (e.g., medical aesthetics, finance) enable specialized checks.
• Task Creation & Auditing: The skill processes inputs via scripts. For local videos, it uploads the file to obtain a video_id before creating an audit task. Core auditing involves text-layer substantive inspection (analyzing voiceover content), a visual-audio checklist (based on multimodal extraction), and industry-specific pun detection. The process runs asynchronously with polling to avoid blocking; multimodal audits typically take 5–10 minutes.
• Result Delivery: Upon completion, a structured Markdown report is generated, detailing scoring across sections and an overall verdict. Locally, an empty Excel template (VIDEO_QC_XLSX_FILE) is created for manual review of visual items. The delivery explicitly states actual points deducted and reminds users to manually double-check visual elements.

Scope and Caveats

• File Limits: Local video uploads are capped at 100MB and restricted to common formats (e.g., mp4, mov); exceeding limits or using unsupported formats triggers an upload failure (exit code 9).
• Duration & Asynchrony: Multimodal audits rely on remote processing, so longer durations are normal; users should wait at least 10 minutes before considering exceptions, with status checks via polling (--poll-task).
• Points Confirmation Mechanism: New tasks require upfront confirmation of estimated points (~20) from the user; execution is blocked without confirmation. Uploads themselves incur no points, but task creation deducts points only once upon first successful result retrieval.
• Technical Boundaries: The skill does not parse video frames locally; all multimodal extraction and auditing occur remotely. Without video input, only text auditing (voiceover or subtitles) is supported.
• Output Rigor: Reports are based strictly on script returns and must not be altered. Visual elements like backgrounds or subtitle styling require manual review by the user per a checklist, ensuring coverage of details beyond automated detection.