amazon-rekognition

What it does

Amazon Rekognition for production image and video understanding enables an agent to analyze existing images and videos to extract metadata, screen user-generated content, build searchable media libraries, and support media operations. It covers labels, objects, scenes, OCR, moderation, image properties, Custom Labels or moderation adapters, stored-video analysis, conditional streaming-video workflows for existing eligible accounts, searchable media libraries, confidence evaluation, S3/IAM/event architecture, privacy, biometric consent, cost control, lifecycle management, and QA. It explicitly states that it should not be used for generating images, editing pixels, creating synthetic media, or judging ethical usability for generation.

How it works

The skill treats Rekognition as an analysis system returning probabilistic metadata rather than definitive judgments. It details image analysis capabilities including DetectLabels, DetectText, DetectModerationLabels, DetectProtectiveEquipment, RecognizeCelebrities, and face-related operations (DetectFaces, CompareFaces, and collections) with biometric workflow caveats. For stored video analysis, it describes Start* and Get* operations to process labels, moderation, text, celebrities, faces, person tracking, face search, and segment detection, with delivery via SNS and retrieval of results. It notes streaming/bulk image analysis are no longer available to new customers and suggests stored-video analysis or event-driven pipelines for near-real-time needs. It covers customization choices between general labels and Custom Labels, and explains moderation adapters for domain-specific improvements. It describes a searchable library architecture involving S3 ingestion, event triggers, synchronous and asynchronous Rekognition calls, metadata normalization, and indexing in OpenSearch or databases. It emphasizes that Rekognition returns confidence scores and that outputs must be evaluated with a defined policy, thresholds, and validation datasets.

When to use it

Use Rekognition when the user asks for: image labels, objects, scenes, OCR in images or stored video, moderation or brand-safety screening, Custom Labels or moderation adapters, stored-video analysis, searchable media libraries, AWS-native async pipelines with S3/IAM/SNS/SQS/Lambda/Kinesis, and face liveness or face search in explicit identity workflows. It also provides guidance on preferred production patterns for new designs and cautions that certain features (Streaming Video, Bulk Image Analysis) are not available to new customers.

What it can touch

This section is included when tools or scripts are specified in FACTS. The skill lists declared tools: claude-code, codex, copilot, cursor. It also references AWS services (S3, IAM, SNS, SQS, Lambda, Kinesis, CloudTrail) as part of architecture and workflows. It does not specify file-level touch permissions beyond those integrations.

Caveats

The skill states not to use Rekognition for image or video generation, editing, or synthetic media, or to judge ethical usability for generation. It notes that streaming video and bulk image analysis are no longer available to new customers, with existing access for those who used the features within the previous 12 months. It describes privacy and biometric safeguards for face workflows, including consent, bounded purpose, retention policies, and deletion procedures. It emphasizes that confidence scores are probabilistic and should be evaluated in context with defined policies and thresholds. License is MIT.