Video Realism Retoucher

The Problem to Solve

Current AI video generation excels at creativity but often falls short in physical realism. This manifests as a range of subtle yet immersion-breaking "AI artifacts": unnaturally cast lighting, repetitive or illogical material textures, harsh color gradations, or incoherent details around subjects. While individually minor, these issues collectively make videos feel artificial. Users are typically satisfied with the original shot, subject, rhythm, and audio direction, and seek a targeted fix for the single most noticeable flaw, not a full regeneration.

How the Skill Works

This skill provides a focused and controlled retouching workflow governed by a "one-shot, one-issue" principle.

1. Freezing the Single Retouching Goal: The workflow begins with precise requirement alignment. The system analyzes the source video and user description to confirm the singular problem to address (e.g., "unnatural shadows on the character's hand"), essential subject details to preserve, source aspect ratio, duration, audio, and delivery purpose. It then presents the retouching plan, model, output settings, and real-time pricing. A clear user confirmation is required before any paid work commences.

2. Query and Execution: Before execution, available video_edit models and their current pricing are queried via beatra.models.list. The edit is then submitted using beatra.videos.edit. All remote operations strictly use the bundled scripts/mcp_client.py, employing a unique client_request_id for each approved logic request to ensure precise task tracking and billing.

3. Tracking and Delivery: Post-submission, the task is polled using beatra.tasks.get until it reaches a terminal state. The final report includes the resolved auto model selection (if no model was specified), the actual output duration, MIME type, usage metrics, and billing facts. In case of insufficient_balance, the response is relayed with the top-up link; the same client_request_id is only retried after the user confirms a recharge.

Scope, Boundaries, and Key Notes

• Retouching, Not Generation: This skill is exclusively for local optimization of existing AI-generated video. It is not intended for multi-shot editing, video upscaling, or content extension. Such needs should be routed to beatra-ai-video-studio.
• Strict Workflow Isolation: To ensure accurate billing and tracking, changing any element of the request—source video, retouching target, model, canvas, or audio settings—constitutes a new request and requires separate user approval.
• Client Exclusivity: All remote communication is performed solely via the bundled mcp_client.py script. Configuring or directly calling the host Beatra Connector, or using REST/OpenAPI fallbacks, is prohibited.
• Recovery and State: In cases of lost task context or expired upload authorization, recovery should first attempt beatra.tasks.list followed by beatra.tasks.get. Creating new paid video edits to recover context is not authorized. beatra.tasks.cancel is only used at explicit user request.