Azure Document Intelligence for TypeScript

The skill exposes the data flows for API keys, DefaultAzureCredential, managed identity, endpoints, document URLs, and training-data SAS URLs, and it recommends managed identity; no credential theft, covert exfiltration, or destructive default is evident. However, it does not state that secrets must not be hard-coded, explain handling or retention for potentially sensitive personal documents, or provide confirmation, least-privilege, rollback, or external-impact controls, so points are deducted.

Installation, authentication, analysis, polling, pagination, and error-checking patterns are broadly coherent, with isUnexpected and long-running polling used repeatedly. Still, the supplied material contains no skill-specific tests or reproducible results; some examples such as receipt analysis omit error checking, and empty results, timeouts, service errors, version drift, and input validation are not covered. Under the static ceiling, the score cannot exceed 10, so points are deducted.

The name, description, and examples clearly target TypeScript users and cover invoices, receipts, IDs, forms, URL/local-file inputs, custom models, and classification. Boundaries for non-fit cases, service-version requirements, input/output contracts, semantic triggers, and Chinese-language support are not stated. The core function also depends on Azure services, with no mainland-China reachability or regional limitation guidance, so points are deducted.

The SKILL.md is readable and progressively organized around installation, authentication, task examples, models, polling, types, and best practices. It includes MIT licensing, Microsoft authorship, version 1.0.0, and the package name in metadata. It lacks a pinned SDK/API version, changelog, explicit maintenance owner and update path, FAQ, troubleshooting, known limitations, and complete prerequisites, so points are deducted.

The examples are directly reusable TypeScript starting points and cover common prebuilt models, URL/file input, custom models, classification, and pagination, providing clear value over starting manually. However, no executed evidence is supplied; output shapes and service-version compatibility are unverified, while secure document handling, retries, timeouts, confidence thresholds, and production integration remain for the user to add. Under the static ceiling, the score cannot exceed 7, so points are deducted.

The skill provides auditable code, package names, environment variables, and one official documentation reference; the repository materials also show generic CI and test infrastructure. They do not provide skill-specific tests, execution logs, pinned versions, or independent corroboration, and this review did not execute the code. Under the static ceiling, the score cannot exceed 5, so points are deducted.